 Good morning everyone. My name is Jimmy Chen and I'm the managing director of StorageX and we have a fantastic symposia for you today around a very important topic, thermal runaway in batteries. With a fantastic duo giving you a perspective from both the industry and also from some of the latest modeling developments that we now have available to tackle such complex phenomena like a thermal runway in batteries. So with that I'd like to introduce our first speaker Dr. Troy Hayes. Dr. Troy Hayes is a practice director of Asia offices and principal engineer at exponent. Dr. Hayes has extensive experience solving complex technical problems in a variety of industries including consumer electronics, consumer products, medical devices, vehicles, industrial equipment and technology product and development. He has conducted technology due diligence reviews of emerging technologies and intellectual property for clients such as investment firms, IP owners and companies interested in purchasing IP or its associated products. He specializes in the mechanical behavior, degradation and failure of materials including metal, polymer, ceramic and glasses. Dr. Hayes has worked extensively on projects involving battery design, manufacturing and failure analysis. So with that I'd like to have Dr. Hayes. Thank you Jimmy and good morning everyone. So as Jimmy said we're going to be talking about thermal runaway in liquid mine batteries and my part of the talk we're going to go through some analyses and case studies related to that and how it impacts the design and things that integrators have to think about when they're designing products that use liquid mine batteries. So as Jimmy said I'm the practice director of materials and corrosion engineering at exponent. I'm also director of our Asia offices. I've been an exponent for almost 19 years. My background is in mechanical engineering and material science. I want to give a special shout out to Dr. Nick Fianza who helped a lot putting together these slides. Probably couldn't have got them prepared in time without him. So thank you Nick. So exponent is an engineering consulting company started back in 1967 out of a group of professors here at Stanford like many many startup companies especially in the big area. We have over 950 consulting staff around the world at this point and over 90 technical disciplines but basically any technical discipline you can imagine. We started as the brainchild for Dr. Alan Tettelman who is many of you if you're in material science may recognize the textbook here on the screen. Barrett Nixon Tettelman three very esteemed professors. So as I said about every technical discipline you can imagine all the engineering sciences human sciences human factors all different aspects of any technical problem you can imagine. We're talking about batteries the groups that we have that specialize in these areas. We have over 80 consultants who spend most of their time working on primarily lithium-ion battery projects and of course that would include vehicle engineering with the proliferation of electric vehicles thermal sciences to study the the heat transfer issues and how fires spread polymer science material chemistry looking at the molecular level how batteries operate and degradation mechanisms mechanical engineering to evaluate how those systems are put together and behave in use as I said I'm in the materials and corrosion engineering practice so looking at all the material aspects of batteries and manufacturing and of course electrical engineering computer science how all these systems are integrated and how you're treating the battery on charging discharging things like that and of course data sciences is a huge part especially as you're getting larger and larger batteries you're getting more and more data from the field that needs to be processed and analyzed. So here at exponent we have with this multidisciplinary approach we're able to evaluate any type of battery issue that that could be out there ranging in scale and complexity from implantable medical devices, consumer electronics, electric vehicles and as we're going up in scale also the largest systems on earth with utility grade storage systems. We have about 25 offices around the US several in Europe and three in Asia and by the little batteries you can see around those are the offices where we have people who specialize in batteries so we really have global coverage. What I'll be talking about first is looking at what thermo on the way is what that looks like in practice and then kind of consider how that influences designers and integrators when you have to think about those things when you're putting a battery pack together. There are three main form factors of a battery cell. Differencing a battery cell and a battery cell is the individual unit that's holding the electrical chemical energy and a battery is that those battery cells combined with an enclosure and protection electronics things like that. So the three major form factors are cylindrical where you have a wound electrode that's put into a cylindrical can. In this case the negative is attached to the can so the very top is positive and the whole can is negative. You have prismatic which is a long chocolate bar form factor cell in this case mostly it's a hard aluminum can that is connected to the positive and then the negative has a feed through on one side and then you have what are called pouch cells and that you have literally like a vacuum seal aluminum bag has polypropylene on the inside nylon on the outside and then you seal that around a stacked or wound cell. In this case the pouch is floating so it's not connected to the positive or the negative and so each of these form factors have different considerations but there's a lot of commonalities as you go through when you're trying to understand those. So what is thermo runway? Well first I can kind of show you what it looks like on inside the cell and the rate of heat generation during those chemical reactions exceeds the ability to draw heat away from those cells away from that that reaction and so then you get a self-propagating thermo on a reaction. So mechanistically inside of a ion battery the main components are starting from the bottom to the top you have the cathode or positive electrode which is a metal oxide painted on aluminum foil usually then the next in this stack you have the separator which is a polymeric porous membrane it allows lithium ions to transfer back and forth through that in the medium of the electrolyte which is between and around all of these components and then on the top of this image you have the negative electrode or the anode which is typically a graphite type material painted on to a copper foil. Now when you start to have thermo runaway or just before you have thermo runaway you'll have breakdown of the protective layer on the surface of the negative electrode known as the solid electrolyte interface and that's a protective layer that's formed during the formation process during battery manufacturing. That's the first thing to break down and start to have some extra thermal reactions but when you get a little higher temperature and you haven't kind of stopped anything from from propagating then you'll start to get major reactions happening revox station and heat generation at your cathode that will interact with the electrolyte and you get to the point where you start boiling the electrolyte you get massive gas generation and that will then cause venting as you saw in the video and once once you vent there's a high likelihood you'll have ignition because you have organic solvents that combines with the air once it escapes the cell and you get flaming combustion so in one way characterized this you can use an accelerator rate calorimeter which is where you have a large adiabatic chamber you put the cell inside and you slowly heat that chamber up like five degree increments for example this jagged line this sawtooth is the chamber temperature the sidewall and you just wait you hold at that new temperature and see if you get anything happening at the cell and so you can very slowly go up until you start to see some exothermic reactions and then once you get to 160-180 degrees this first exothermic is probably associated with the anode but then once you start to get the cathode going off you get this vertical line where you just get a rapid increase in temperature you saw in the now penetration test that just happened very quickly your voltage which is the green line drops off as you start to get shorts within the cell and the temperature this is the external cell temperature gets above 400 degrees C and that's that's what happens so what happens inside the cell well the temperature outside gets to like I said it could be as high as 400 or a little higher degrees centigrade inside the cell it gets much hotter so typically it gets to about six or 800 degrees C inside the cell well if you look at the the two metals that are inside the aluminum that's for the positive electrode and the copper for negative current collector um what I have here is a part of the phase diagram so we have zero percent copper on the left up to 60 percent the melting temperature of copper is about 1,085 degrees centigrade so when thermal runaway happens you get to six or 800 degrees it's not going to melt copper but it will melt aluminum which melts at 660 degrees C well there's two different ways you can actually get melting of the copper in the cell one is as soon as you melt the aluminum starts floating throughout the cell it will alloy with the copper and then you can get a eutectic alloy that can melt as low as 548 degrees C and so you get a lot of melting of the copper winnet alloys in addition if you have a short circuit where your i-square are heating from a short between the positive mega electrode which can happen during the thermal runaway process you can get an excess of this type of temperature looking at those when you open up they look substantially different the alloying reaction you can actually see blobs of aluminum sitting on the copper and you'll see holes like almost cookie cutter style holes where those have fallen away if you look on an scm in backscattered mode you'll see often a halo but where you have a lower density alloy around these holes alum that where you have the melting of the copper on the other hand you have these beaded edges where you have the the metal wants to minimize surface energy so it'll it'll beat up around those edges and that's what it looks like with pure melted copper so what do cells look like after thermal runaway well it really depends on a few different things and in these slides we're going to look at two different aspects here one is state of charge and one is the the age of the cell and so here this is a test we did on some cells from 2013 here at 50 state of charge it looks almost like a brand new cell you can't see much you see some winding movement to the right at 100 you can see a lot of damage but no obvious bright spots what they would expect in in melting which they look bright in a CT scan now if you overcharge the cell and that's where you can start to probably get lithium plating in the cell then you'll start to see more melting because lithium burns very very hot and you can melt the copper so you'll see a lot more bright spots all right so this is what a cell from 2013 might look like in thermal runaway at different states of charge if you unroll the cell at 100 state of charge you find no melting whatsoever there's some alloying spots as similar to what we talked about before but no copper melting so there was no direct short circuit between electrodes that had a high enough current density to to melt copper compared to the overcharge overcharge you see a lot of melting where you have lithium plating you had shorts with much higher energy okay so if you knew you had a cell that was 2013 time frame and you you were evaluating something forensically after a fire you might be able to make some determinations about the state of charge but compare that to 2018 and newer cells where you're going to higher voltage higher energy densities here's a cell that we forced in the thermal runaway at 100 state of charge and you see a lot of melting throughout that cell drop it to 60 and then you start to see damage and some alloying but but no melting so equivalently the 60 state of charge in some newer cells looks very similar to the 100 state of charge in the 2013 type cells if we open those up indeed when you look at 100 state of charge you'll see the the massive copper melting as you saw in the overcharge case until 2013 so it's important to understand you're getting more and more energetic failures with newer and newer cells also depends on the actual history usage history of the cell and what what it's seen in the field so what are the other considerations and why is state of charge SOC so important well obviously the state of charge has to do with whether it's a discharge cell or a charge cell so naturally if you have a fully charged cell you put more energy in you're going to get more energy out that's true but it's not just that simple so if you look at this plot there's three different lines so the one interesting fact is on the bottom you have mass loss which is on the right axis here and you can see whether zero percent or 100 state of charge the mass loss really didn't change now that's only true in this case it's a study that was on pouch cells so in pouch cells remember you had that vacuum seal bag it has nine uh nylon the outside polypropylene on the inside those seals are made by literally melting the polypropylene on those three seams together when you close that vacuum bag well when you go to thermon away on that type of cell you don't get a very concentrated vent because all that polypropylene will melt very easily and so you get a diffuse vent around the three at least three sides of the cell as a result you don't get a very strong jet of venting gas and so you don't get a lot of mass difference with state of charge they would look much different with a prismatic or a cylindrical cell where you have a directed vent but another interesting thing if you look at the total heat release which is this turquoise curve you do see a difference between zero percent 100 there's not that much you may be about 40 watt hours at zero percent maybe 60 at 100 so it's a difference about 50 but what really makes a difference and you see this dramatically on the blue line which is a left axis is the heat release rate how quickly the energy comes out so at zero percent the amount of heat generation is very low and then it's very fast when you have a 100 state of charge and so this has a lot of implications for propagation and whether that will actually be something you can suck the heat out and try to stop through heat sinker or different mitigation measures so of course when you have therm runaway the cell that failed will of course be totally dead and it risks igniting flammables around if it's an individual cell or something you have you could possibly start a fire and then of course the bigger risk as you're getting to larger multi-cell devices is that that becomes a propagating event where you start to cause more and more cells to go off and as we'll see in a little bit that can happen and gets very energetic so how a cell actually goes to therm runaway it really depends on the chemistry whether using some of the higher energy density chemistries or you know LFP is considered safer you have NMC and various different metal oxides that are used in the cathode depends on the cell design whether it's prismatic a pouch or a cylindrical and as we just talked about the state of charge also depends on the initiation mechanism whether it the therm runaway happens from an internal short an external short from external heat attack so all those things will affect the behavior during therm runaway one thing that's really important to take home from from this lecture of nothing else is therm runaway is really stochastic so even if you have nominally the same exact setup and you do the same test one two three times you might get three different results so it's really important when you're thinking about if I'm going to do a 9540a or some ul test on a huge system and it doesn't have a catastrophic failure you can't just pat yourself on the back because if you do two or three tests you might see very different results and we're going to talk about why some of those things are so when you take these individual cells and put them into a battery pack obviously one of the goals is to get the highest energy density possible and if you have cylindrical cells the highest densing density you can get is through a closed pack type morphology so you can put all these cylindrical cells in a closed pack arrangement and put them in a large pack weld all of them together but you're kind of conflicted because if you put all these together if you do a therm runaway then you get very easy propagation one of the ways to prevent that is to have more spacing between the cells possibly put some material between them to insulate them or heatsink them and so you're really battling the energy density and and some of the safety and management type issues so what happens when you have a therm runaway in one of these cells in a in a large cluster well the first cell that goes off you're starting at ambient temperature so it takes a lot of its own energy to heat up and at that time it's also heating its nearest neighbor cells if that if propagation occurs if you're not able to stop that reaction from heating to the second cell when the second cell starts to go off now you have proximal heating not only from the second cell oops but also the first cell and so the the cell is adjacent to the second cell will start to heat up even more rapidly because they've been preheated by the first cell and now they're getting heating from the second cell so this starts to happen more and more rapidly and you get a really fast reaction and a lot of energy release so going back to the stochasticity of this here are two nail penetration tests on normally the same design and so the first one is the same video we saw before there is gas coming out with the nailers but most of the gas came out at the cap of the vent mechanism that's where it's designed to release its its material based on the the vent disc that it has built in the second test though looks like that's kind of pausing a little bit but basically no venting at all happened at the cap in this one instead all of it came out the side so you had a side rupture in fact we did many of these tests and here's a three for example the first one is the first test you saw and it vented out the top which is where it's designed to release that so these cylindrical cells have a burst disc if if you have over pressurization it's made to come out at that end the second two they all vented adjacent to or or nearby the nail pen and so obviously the fact that you penetrate it with a nail has a huge influence on that but the fact is that these tests were run in the same way and you get a different type of of a failure mode now going back to that the state of charge and how you saw you didn't see much mass loss difference between the different cells in a pouch cell design that's also you can see consistent with this test where the the highest mass loss was actually in the cell that vented as design and so mass loss is not necessarily a bad thing if you can if you have a way of channeling those gases and channeling that ejecta away from the other cells then your total heat capacity and can be managed and so you can hopefully prevent that from propagating to other cells so looking at this there's a diagram if the cell vents as designed this the vent gases come out the top you can imagine that might be easier to manage than if it starts to vent into its nearest neighbors and if you start to have this side venting which can happen for a number of reasons for example when you have an external short that the thermal one-way reaction itself actually happens when that drives an internal short that internal short location from that external short can occur at any location within the cell well if it happens to occur near the edge of a cell then that'll weaken the can it'll reduce the the yield stress and so when you get ever pressurization inside that location might be weaker than the top vent and result in side venting if you imagine similarly if one cell side vents it's going to attack the other cells on the side weaken their sides and could also result in additional side venting and so these things are influenced by the cell design whether you have one vent two vent how thick the can walls are but also how you're how you're working managing the heat if you do have propagation thermal one-way in one cell how you manage that the other cells and their exposure to that heat also just dependent on the the age and state of charge of the cell so in this particular case we initiated a cell in the middle of a cluster it experienced a side venting you can see here the initiator vented out the cap and the side because of this four of the six neighboring cells also had side vents so again not hard to imagine if you basically put a cutting torch and blow into the side of a cell it's going to vent out that side rather than out at its top where it's designed to do so then how do you use this information to design a battery pack well there are a few different mitigation strategies thermal is one so in the pack in the upper right you can see you have an array of cylindrical cells and then you have this pink material which is basically that the annular region between those closed pack pattern of cells is filled with this foam so in this particular design the thermal strategy was to insulate the cells and if you can insulate the cells well enough then you can prevent the heat transfer from one cell to going through runaway to its neighbor to the point that hopefully it won't it won't propagate and in these large packs they're often designed such that if one cell experiences some runaway the pack which has a whole bunch of cells in parallel you can see all these have the same orientation with the positive pointing down so these are all at the same vultures are all tied together in one one bank and so these are designed that if you lose one cell it's able to continue to operate you also can do the opposite you can use heat sinking you sometimes people put phase change materials between the cells to absorb that heat if you have a thermal reaction there have been packs that use liquid cooling that is channeled through the annular region between these cylindrical cells also you can have forced air cooling so in the cylindrical design it gives you a lot of options that way if you go to more of a prismatic or a pouch cell design it's you have fewer options because again you're trying to max out his energy density and so you stack all these up to use all the space as you can see in the lower right but if you don't have barriers between and space between those cells then it can be very difficult to prevent propagation so in this one on the lower right they actually use some kind of a urethane to so once the pack was made they actually in in case the whole thing in a in a urethane to try to to manage the therm one away you can also manage propagation electrically so in the lower left you can see you have these small wires connected to the bus bars so that all these cells that are connected in parallel if you were to have an internal short in one cell what happens is because all the cells at the same voltage the cells that haven't been shorted will all dump their energy into the cell that that suffers the short circuit and so if you don't have if you have a large bus bar connecting them then all of that energy can be dumped into the cell that has a short circuit and it's more likely to experience therm one away but if you have these small fuse type designs on these ligaments then that current when one cell experiences short circuit will be cut off as these links will fuse open so similar to the we talked about venting in cylindrical cells each of the tops of these will have a vent design you also need to accommodate that mechanically in the design of your pack because as you start to have cells that go to therm one away and maybe propagate they're going to generate a large volume of gas and so you need to be able to make sure that doesn't become an explosive event at the pack level and so you also need to develop generate mechanical vents at the at the pack level to accommodate those types of things so the first line of defense against propagation and therm one-way is to prevent therm one-way at all from happening and you can do that through cell design and pack design so but in addition to that you also can do that through how you manage that in the use of the battery pack so with you remembering back to when I said 2013 versus 2018 cells you can actually get the 2018 cells to behave just like the 2013 cells if you back off the state of charge and so that's what some electric vehicle manufacturers will do is that they actually won't use the full capacity of the cell if you only charge to 70 75% state of charge then any reactions that could cause them run away in an individual cell are much easier to manage and your cycle life can improve from you know something that might be 500 to 1,000 cycles can get you easily to 10,000 cycles super important considerations in pack design cell selection obviously you want to use top tier manufacturers you need spacing between the cells again it's a battle between energy density and and having that spacing to allow you to to manage the thermal heat transfer between them vent gas management how are you moving that gas away from the cells and away from the pack whether you need thermal barriers between cells and different banks of cells or whether you want to actually heat sink them obviously electrical shutoffs we talked about you can do that at individual cell levels you can do that at block levels and at pack levels and then mechanical protection obviously when you get into any devices that are going to be dropped when you get to electric vehicles where you have crashes you need to make sure that you don't impinge on the cells and cause these kinds of catastrophic reactions once you have your design and it's in the field it's super important that you start to gather data and this becomes more and more important the larger the system you're dealing with we've had we've investigated fires where there are literally thousands and thousands of cells involved in a fire and the traditional methodology of taking the initiating cell identifying it out of a group of you know maybe five cells or out of 15 cells it's not so challenging but if you're trying to identify the initiator by looking at every single cell among thousands of cells it's really challenging especially the fire has been so great that they're now a whole bunch of just electrodes and junk all over the floor it's very hard to reconstruct that so if you're able to look at the data real time then you can start to pick out where the the incident may have started and if you're monitoring the correct things you're monitoring voltage of each bank the capacity of the different banks temperature at various locations and and the charge efficiency those types of things then you can start to manage that and identify early before you get a thermon away you can start to identify some markers and manage those systems and through that you can limit the the state of charge if you're starting to see some stresses going on you can reduce the state of charge at full charge you can also deactivate certain modules to help limit your risk and so again data management now that we were able to monitor and control big data the more data you can you can take from your system the better because you're able to identify trends and possibly put in algorithms that will will identify issues before they actually start but super important that the quality of that data is good and that you're monitoring not only batteries after they fail but actually when they're in the field so you can hopefully identify these and and pull them out or get them shut down before they have any catastrophic reactions also very important to compare the the performance of of different packs veteran operation so how frequent do you have thermon away well the numbers that are thrown around is you know it's about one in 10 million to one in 40 million cells at a cell level if you have a top tier manufacturer so it's super rare but it's still not something that is perfect even though these have been made since the mid 90s it's still not a commodity they still have these catastrophic failures and then if you think about a car where you have thousands of cells in fact in the Tesla Model 3 with an 82 kilowatt hour pack you have 46 hundred cells it doesn't take much math to figure out you're going to get this kind of catastrophic reaction in about one in 2000 to one in 5000 cars and so if the whole car burns up out of every two or 5000 cars that's just not acceptable and so you really need to design for cell failures you need to assume that cells will fail we're going to try to prevent it we're going to do everything we can to prevent that from happening but we also need to design a system that either through heat sinking or thermal management won't let the cell go thermal or protect that thermal management so that if it does go thermal that it won't propagate to the neighboring cells and create a large chain reaction wrapping up here before Alania it will we'll go into talk shortly about modeling it is important to take the data that we've learned through accelerator rate calorimetry through various different tests that you do and and feed that back and start to very early on in your pack design model how your thermal reactions happen and so you can start with reduced order models to simplify that problem and evaluate how you get cell to cell propagation and how that propagation can then move from a small problem to a large problem and that's going to be the focus of Alania's talk which I'm very excited to listen to here shortly but wrapping things up here key takeaways preventing thermal runaway it's not trivial it's not just a matter of making sure you have the best manufacturing process in the world really takes a lot of thinking a lot of design mechanical electrical thermal in addition to best manufacturing practices you have to design a battery pack with the assumption you will have failures of cells and you have to do everything you can to then try to prevent those from becoming a larger failure especially on systems where you're designing you're integrating hundreds or even thousands of cells and finally testing under various different expected use and abuse conditions for new and new cells can really provide a lot of great data that you can use then when you're monitoring real-time data in your system and help designers and integrators then use that to develop battery packs with that I want to again thank Nick Fanza for helping to put these slides together and thank you for for listening. Fantastic Troy thank you very much for that survey of thermal runaway in your case studies so I I want to just kick off with a few general questions so most of the examples that you described were in the context of electric vehicles is that is that accurate or do you I mean I think it was a pretty pretty notable example before of consumer electron consumer electronics with iPhones and so on so I'm just wanting to comment a little bit about sort of the different applications and thermal runaway. Yeah I mean certainly I would say we cut our teeth on thermal runaways of batteries and consumer electronics and starting there and as I said we work on all types of scales so my personal projects I have literally I'm working on a battery that is implanted in the skull of an implantable medical device I'm also working right now on a grid scale storage and so really the same the same considerations apply to all scales of systems in one case you're talking about how does heat transfer to bone and to the the matter inside of the the the skull cavity in another case you're talking about how you're transferring energy through these these giant banks of cells but ultimately the process of thermal runaway and the management of those thermal issues is is equally important in the different cases obviously the big scale systems release a lot more energy and so they they tend to be more on the news. Yeah so in that context then are you seeing any are the different types of cells prismatic pouch cylindrical do you see any differences in thermal runaway potential of those formats? Yeah certainly the pouch cells so if you look at the different electric vehicle manufacturers there there isn't a standard form factor that's chosen Tesla tends to use cylindrical cells but if you look at Ford and GM they're using large pouch cells pouch cells they they definitely release energy differently as we talked about the venting is more diffuse around the edges which can be helpful in certain circumstances but in my experience it's also very hard to to limit heat transfer between cells if you don't have a gap right so the having those annular regions between cylindrical cells provides a natural cavity that you can use for thermal management and so in designs that use pouch cells they really have to figure out how much spacing and do you need to put ceramic between each pouch cell but ultimately anything you've got you're creating those gaps you're reducing the energy density and so it's always a balance you need to find the perfect balance where we have all the thermal management you need but you you have the energy density to be able to you know have your your pickup drive 600 miles or whatever it is. Yeah so there is a well I guess there's a general perception especially when I talk to the public right now that the that the there seems to be a greater number of recalls with electric vehicles and I'm as someone who's probably in this all the time do you perceive that that's the case compared to say internal combustion engines and I'll come back to that a little bit but I'm curious on your assessment. Yeah it's interesting yeah there there certainly been some very high profile recalls we're still learning how to do this right that's that's true and one of the things I think is distinct between an internal combustion engine fire and electric vehicle fire I mean there are still a lot of internal combustion engine fires but there's a couple of differences one when a fire an internal combustion engine happens once the fire is over it's over one of the the keys about the lithium-ion batteries is if you're successful at preventing propagation but you had partial propagation of a battery pack now you've got let's say part of an electric vehicle that has had full thermal runaway but an adjacent module that you're able to stop the reaction before it did maybe through first responders or what have you but now the adjacent module it can be thermally damaged it can have all kinds of water ingress and so there couldn't be a failure that is going to be delayed a matter of hours or days and so there's all kinds of stories in the news where you have an electric vehicle that the fire they put the fire out several hours or days later catches fire again and again and again so that's one difference and another is a lot of the fires with with lithium-ion batteries in general by and large start during charging so when you charge a lithium-ion battery you get expansion of the electrodes and so your wear-through type failures related to stress often happen during charge that couple of a factor highest state of charge where you have the most energy to pour into a fault happens when it's charged well where do you do charging electric vehicles they're at home so a lot of these electric vehicle fires are happening around residential areas as opposed to on a freeway or somewhere where it's where it's less catastrophic so there's a lot of sensitivity to that but just like autonomous driving it's kind of in the spotlight right now and so you know we're trying to do everything we can to to make that right but you know obviously we don't have the experience with as we do with internal combustion at this point yeah that's that's that's fantastic i think um one of the things that uh i maybe you can just um give your thoughts on this what are the common root causes of thermal runaway so you mentioned charging or and then in some cases a collision or something what do you see as sort of the example of root causes that might be related degradation whatever you know what are you what do you see are the mechanisms for the thermal runaway um and that also goes back to me to your question about the different uh how different cells behave so the ever cylindrical cell because the can is very robust um even if you were to drop it or have some mechanical stress on the outside it doesn't usually cause a catastrophic failure but if you imagine you just have a soft pouch like you have in your you know whatever cell phone you have in your pocket if you do something to crush that the cell is is a soft vacuum packed bag and so what we've seen is in cells that that are pouch cell design there's a much higher percentage of thermal runaways that are external to the the cell itself so whether that's during pack assembly or during use because it has a soft bag um it's much easier to cause a failure in fact if you look at the Samsung Galaxy Note 7 recall one of the two failure modes was basically the enclosure in the phone didn't have enough room for expansion electrodes and because it didn't have a hard case that basically impinged on the edge of the windings so on pouches uh external mechanical environment is super important um but regardless of design there are a lot of different commonalities one is you know thermal obviously you can have external heat attack um and when you look at fires uh lithium ion batteries because it's very well known in the news they can catch fire then we find they're quickly pointed to in in fires because they're a hotspot right even if you hadn't regardless of start if you have a electric vehicle in your garage and the fire propagates there that's going to be a super hot spot and so um that certainly thermal can set things off degradation you know one of the super important things is the electrode balance between your positive and negative electrode you need to have more capacity in your anode your negative electrode than your cathode so when you fully charge you don't get plated lithium on the surface of your anode um and all the cell manufacturers know that but one of the things that can happen is if your electrodes age um disproportionately and if your anode ages more quickly than your cathode all of a sudden you're you can't accommodate all of the the lithium during charge and you can get plating even on a cell that was balanced well originally uh lithium plating also can be exacerbated if you have you know charging at low temperatures um if you're charging at too high a rate um charging a cell maybe you could do a one c uh you know charge in an hour on a cell when it was new but because of degradation um your lithium transport is somehow inhibited and can result in lithium plating so lithium plating is a common cause of uh of failures you can get dendritic growths and things like that but uh yeah there's a lot of different failure modes electrical if you don't manage your charge algorithm properly um if you you don't manage discharge um and of course mechanical it's quite a few different reasons and then of course internal manufacturing uh can also some of the recalls are talking about electric vehicles the the battery manufacturers pointed to manufacturing issues on the on the line that cause those failures so certainly they can they can cause failures as well okay so um I imagine from your comments and that uh this uh transition potentially to fast charging would need to be done carefully uh since that the the charging rate could play a role in this yeah no absolutely um because again it's not a constant target you you have a fast charge rate that you verified on a brand new cell is a cell with you know 10 000 cycles still going to be able to accommodate that fast fast rate of charge and so then it goes back to what I was talking about with data analysis if you're able to characterize for example the the impedance of the cell and the charge efficiency then you can start to suss out some of those characteristics that you know need to reach as you be at a certain minimum level in order to accommodate that fast charge so it might be your fast charging will be able to get you know what do they say 80 percent in 10 minutes when it's brand new but maybe if it's a five-year-old car maybe it's 80 percent in 20 minutes so it's really important that you consider that with the lifetime of the cell do do you believe right now that there are adequate sensors that could give you that kind of insight that you might have potential issues that that are in place right now where you can fly in a potentially place a particular pack or module or or as as a program yeah I mean there certainly sensors exist and the measurements exist but it's always just like your energy density it's a balance are you going to put in if you have 10 000 cells in your in your car well it'd be great if you had 10 000 thermocouples and if you were able to measure obviously they're in parallel so you can't measure individual impedances but you know there's a limit on how much you're going to put into measurement on test packs you certainly could but you need to balance that and so you need to find the sweet spot of finding those key locations and and what data you need but but you know we're learning as well and as manufacturers are kind of learning what is the most important measurement then they're putting that into the the firmware and and updating these algorithms so they're more and more accurate at detecting failures once a failure happens often that's a learning process and so then they could feed that back and say okay now we're looking at the data look at this we see this spike just before failure can we then implement that into our you know systems are still in the field to say if you start to see that ramp up then we back off the charge or we disable the pack yeah so a really sort of high level question to to end it you know we're more watching down this this path of electrification and where now we see energy storage devices at not just you know mobile stationary electric vehicles across the board potentially even considering energy storage in residential to offset you know disruptions how do you see or how would you recommend you know there's there's a big concern as we start down this path that there are things we're learning as you put it that we're still learning right and so a general general question long you know how how would you orchestrate this so that we have a more smooth transition and minimize or leverage our learnings across the various use cases it seems like applications a big deal here yeah well it's interesting because you obviously if you had everyone adopting the same form factor in the same battery pack we could do a lot to to make that efficient but ultimately I think you don't want to put all your eggs in one basket and that goes for clean energy it goes for for batteries and so really one of the things that's super important and it's that's not really the topic of this this talk but you know how are we making this green really green if you're driving your Tesla and you feel like I've got an EV well if you're using fossil fuels to charge it how green is that if you didn't extract lithium in a way that's good for the environment then you're just kind of moving the the the bad effects away from your neighborhood and so that whole process and they're starting to have you know you can have QR codes that looks to make sure you have basically like conflict-free lithium and these things are going to be more and more important to make sure that as a whole as a world we're actually improving through this technology and so that's really it has to be addressed from all angles and I think lithium-ion batteries is currently the big bet that the world's made you know with fuel cells and other things I think that you know we shouldn't just discount and say okay lithium is what it is we need to keep building and look for non-organic electrolytes and and keep pushing forward but but don't forget about these other things where we might be able to make a big difference. Okay fantastic well thank you Troy I thank you for the presentations and discussion and I'd like to now introduce our second speaker and we'll bring you back Troy after we have the presentation by Lenya. All right thank you thank you Jimmy. Thank you. Okay our second speaker is Lenya Batiato who's an associate professor here at Stanford in the energy science and engineering as part of the new door school sustainability. Dr. Batiato's research focuses on understanding modeling and predicting complex multi-scale multi-physics systems with cross-cutting applications in the energy landscape ranging from electrochemical storage to CO2 sequestration and hydrogen storage in the subsurface. She uses a combination of rigorous mathematical theories numeric and symbolic computing to develop advanced multi-scale multi-physics model. She attained an MS in environmental engineering with highest honors from the paleotechnical Milano Italy in 2005. She subsequently obtained an MS in engineering physics from the mechanical and aerospace engineering department of the University of California San Diego and she completed her PhD in 2010 at UCSD also. She received the DOE young investigator award in basic energy science for innovative work on multi-scale models and course media. With that welcome Lenya. I'm Lenya Batiato. I lead the multi-scale physics and energy system group in the department of energy science and engineering at Stanford and today I'm going to talk about some of the work that we have recently done in the context of thermal runaway and in particular about upscaling and automation and how we can use and advance our understanding and modeling of multi-scale of thermal runaway through symbolic and numeric computing. And I know this is a lot. It's a mouthful. I would like to emphasize just three concepts from that title. Thermal runaway, multi-scale modeling and symbolic and numerical computing. So thanks to Troy. We now know everything we need to know about thermal runaway. So thank you Troy. I in this talk I'm going to focus on multi-scale modeling and symbolic and numerical computing and the claim of this talk will be that battery systems present unique modeling challenges because they are multi-scale and multi-physics systems. However, we can capitalize on their very multi-scale nature through novel technologies in symbolic and numerical computing to achieve predictive accuracy while not compromising on computational costs. And I want to link back quickly to one slide that Troy showed towards the end where essentially he showed numerical simulations of thermal runaway in a group of cells. I think it was about 60 and the importance of developing compute accurate computational models because really computer-aided design can significantly reduce the cost of making design choices while ensuring accuracy and safety. And so really the need is to develop accurate predictive models for computer-aided design. On the top I have here a picture of a Tesla battery pack. You can see that there are over 7,000 cells in this particular design and then here we see a module. It contains over 400 batteries and then this is just a zoomed in view of how we can conceptualize it as containing three domains. The battery cell which is the pink domain, then the white domain which represents the packing material where these batteries essentially are being encased and then a third domain which could be for example cooling types. And we know that thermal runaway as Troy explained can be caused by a number of different causes. One of them is the onset of exothermal reactions that increase the temperature of the battery which speeds up the reaction process and that leads to this catastrophic event where one cell can undergo thermal runaway and then this can propagate down to neighboring cells causing essentially complete destruction of a number of different cells or even the entire pack. And so prevention clearly is critical and from a computer-aided design point of view maybe some of the questions that we would like to be able to address is what are the best geometric parameters for preventing thermal runaway and can for example minimize cooling pipes usage by optimizing unit cell geometry or for example what kind of material should we use for the packing in a way that we can again reduce thermal runaway or the chances of the risk of thermal runaway. Now why are lithium ion batteries multi-scale multi-physics systems? So we refer to a multi-physics system as a system that in which a number of different physical processes occur concurrently and are coupled together. So in this particular case for batteries we know that we may have you know we have lithium ion intercalation the intercalation reaction we have diffusion in the electrolyte in the solid phase we have electrode electrokinetic effects we have heat transfer we can have you know mechanics deformation. Batteries are also severely multi-scale systems from material science we know that a lot of design actually goes into the material design of the electrode particles and this is typically a very fine scale right it's sub micron scale. However the performance of the system that we are really focused on goes up to the meter scale so we're interested in really how this design a very small scale and intermediate design at the cell scale or module scale can really affect the performance of the entire battery pack. So in terms of spatial scales that we have to cover in order to be able to develop models at the system scale it's really technically six or seven orders of magnitude at least. For time things becomes even more complicated because reactions may have characteristic temporal scales of milliseconds while you know aging can occur over months or years and so that leads us easily to seven to ten orders of magnitude in time. So the development of rigorous and predictive system scales model from first principle is highly technical it's time consuming and prone to error but now really what is the specific challenge to model such systems rigorously? So clearly if we start for for example from the module scale we could use a high fidelity model where we would solve you know our model in this very intricate domain. Now the advantage of using fine scale or high fidelity models is that these are you know accurate we have confidence in the models or equations that describe processes at this particular scale but the problem is that they are computationally expensive. I'm assuming that the simulation that Troy showed just for 60 cells must have taken quite a bit of computational time however now we are really thinking about modeling systems that have two orders of magnitude a larger number of cells so we are going from 60 to 7000 and so an option to overcome this issue is to say well we don't really maybe need to use these high fidelity simulations at this level of granularity instead we can try to build equivalent representations or reduced order models or equivalent continuum representations these are I'm going to use all these terms kind of interchangeably where essentially now my very fine scale domain can be represented as equivalent continua where the pink domain will give me information about the average temperature of the battery domain while the white domain will give me information about the average temperature of the packing domain for example. Now the advantage of using these equivalent or representations or upscale the representation or reduced order models is that they are computationally cheap because the domain is much simpler I can resolve it pretty well with lower computational burden however the model themselves so the partial differential equations that we need to employ yourself will be approximate so having control on the approximation error becomes very critical especially because safety is involved and the process of deriving and determining this equivalent representation from their accurate high fidelity models happens can be done rigorously through some mathematical theories that are broadly indicated or represented by upscaling theories so and that's where the the difficulty is so we start from some governing equations that we have confidence on but we really want to determine what are the models or the reduced order models that we can deploy with accuracy at a scale such that we can accuracy but but we can also retain a computational efficiency so the objectives of the work that we are doing is really to to develop a complete multi-scale modeling strategy that includes model development model verification and validation and we would like certain very specific features of this framework we would like to have fast development time we don't want this framework to be developed in there we go I'm sorry in you know 10 years so we wanted fast we want again controlled accuracy for our priority estimates of the error because we want models that are reliable when deployed we want computational efficiency and we want also transferability framework generality what we mean by this is that if we develop a framework for a specific physical process we would like to be able to have the flexibility to generalize to include additional processes as Troy mentioned the reason for thermal runaway can be varied and so we would like to have that flexibility to include additional physics as we improve the generality now the three type of questions that I'm going to address in this talk are what are these continuum scale models what are these reduced order models how do they look like when are these models reliable and what are our strategies when we cannot use them to do computer aided design and so in our group really we look at the entire life cycle of multi-scale model development for different engineering applications again we have the left branch is the physics based model deployment these are strategies where we essentially develop models rigorously mathematically we also identify during these mathematical procedures the applicability conditions under which these models can predict fine scale processes correctly and accurately so these applicability conditions are then turned into what we call diagnosis criteria for the model and we use then this diagnosis criteria for optimal model deployment and for example we would like to understand you know for certain specific application what would be the best model to use and the deployment should be optimal in terms of two parameters accuracy and computational cost and so this is just a snapshot not updated of the people who have we have in the group really we focus generally on multi-scale multiphysics modeling of engineering systems and we have a specific focus and strength in fluid mechanics transfer reactive transport broadly in engineering physics numerical methods algorithm and machine learning and AI I didn't put applied math but that's kind of imply so however today I'm going to talk about the work that two ex-members have done in the context of thermal runaway Dr. Yao is now assistant professor at Texas A&M and Dr. Patrick with Lawrence fellow Lawrence Livermore National Lab and Sishin Ping with currently a student in the group so let's start with the first few questions so what are the continuum scale model equations that reduce order model equations that we should be using when are they reliable and how can symbolic computing help so as I mentioned this translation between high fidelity model and reduced order models can be rigorously performed through upscaling methods these are a class of mathematical theories that can get us what we need so they can help us generate the mathematical models that that we need and they also help us identify the conditions under which the specific model is valid and predicted what which we call applicability condition so the general idea of these models are that let's say that we have a description at a fine scale we identify and or define a volume which we call representative elementary volume and then based on this volume we can define an average in this particular case since we are looking at a thermal runaway the the quantity of interest that psi would be temperature so we can define an average temperature as an integral over this volume and then the general goal is the following so let's say that we have a high fidelity model so the temperature of the of the cell satisfies a certain a certain partial differential equation a certain model then the idea is that you apply this averaging operator to this partial differential equation and then with some math math that is all hidden behind this simple arrow what you're what you want is actually another model that now holds for the average temperature and so once we have this model that's our reduced order model our continuum model then we can solve it numerically it's much cheaper and then we can of course make predictions at large scale and there are different methods that you can use to perform this procedure we generally use homogenization theory but all other methods would give you roughly the same information now the idea then is that okay if we know the physics at this particular scale for example at this cell scale then we can derive our our upscaled models for the for this particular scale and thermal runaway then we can have for example a conduction equation that describes a temperature evolution in the packing material and another conduction equation in the in the cells but now the difference is that we have a heat generation term that Troy was talking about where essentially the heat generation term can be characterized experimentally and the way it looks it's generally something like this where the battery generates some heat it's like low level heat up to a certain temperature but when the temperature of the battery reaches a certain values then the heat generated ramps up so the battery undergoes thermal runaway the temperature continues increasing until you get complete burnout at which case essentially the cell is dead and does not generate any heat and then of course you have technical boundary conditions that represent how these domains are connected thermically now of course as as we mentioned before we can always solve these partial differential equations these models on our complex domain and we with these equations we can both model an ignition scenario and a no ignition scenario so on the left hand you see a no ignition case where we have the center the cell in the center undergo thermal runaway but you see that the heat is propagated through the pack but it does not activate the other cells to go into thermal runaway on the other case instead we have an ignition case where the first cell again fails undergoes thermal runaway but this heat is propagated throughout and triggers thermal runaway and the catastrophic event that in fact affects all the battery pack so the problem with those simulations is that as I mentioned before they are very expensive so we can use these theories these upskilling theories to derive reduced order models and we have everything in place to do that we can do systematic generation of macroscopic differential equation and we can also use this theory and to develop multi-scale models that seamlessly connect adjacent scales so we technically require no parameter no parameter fitting a course scale if you have the appropriate information a finer scale however the problem is that really those mathematical derivations that I hid behind this arrow are incredibly complex they are tedious time consuming prone to error and if we also want to derive the applicability conditions of these models which tell us when these equivalent representations are actually accurate that can take months or years of derivations so and when I say it's bad I really mean it so this is just an example of the results of some of these upscaling theories and the derivations is not present and so our suggestion here is to essentially automate we want to get accurate models at the appropriate scale that are that are cheap to solve computationally but we don't want to have to do these calculations by hand and so we want to use symbolic computation to automate the entire model development of these complex systems so now the concept of using symbolic computation is not new in fact it was first introduced by a woman Ada Augustine Cantes of Lovelace in 1842 and she took she was highly educated that she took notes on the writings of Babbage concerning his analytical engine which is considered the first computer and one of these notes reads as follows so many persons who are not conversant with mathematical studies imagine that because the business of the Babbage's analytical engine is to give its results in numerical notation the nature of its processes must be consequently arithmetical and numerical rather than algebraical and analytical this is an error the engine can arrange and combine its numerical quantities exactly if they were letters or any other general symbols and in fact it might bring out its results in algebraic notation notation where provisions made accordingly and so the idea here was really to use symbolic computing to automate the upscaling procedure symbolic computing is the science and technology that aims at automating a wide range of processes involved in solving problems in mathematical physics and symbolic computing has been used routinely in a number of other branches of applied in theoretical math and computer science but for some reason this tool had never been involved in or used to advance high level manipulations of in mathematical physics and so that's precisely was our point we wanted to automate all these upscaling procedures that are very important to develop rigorous equivalent representations for these more complex processes and so we have done this through a code symbolica this is a symbolic software developed in mathematical which automates a completely the upscaling procedure and so the idea is that again we have a fine scale high fidelity model that can look as ugly as you want then we want to build the model at the continuous scale which we may we may not know how it looks like and again this if we were doing it by hand would take months to years to derive but now we don't want to do that we want to allocate all the computational resources to symbolica and we speed up the process by five orders of magnitude so we now can develop models that are accurate and predicted in seconds symbolic as an underlying algorithmic structure that i'm not going to discuss but essentially if you check symbolica in action for the thermal runaway problem it provides you with these reduced order model in 62 seconds now i'm not going to discuss the pd's that are on this page but i want to just highlight that these pd's these reduced order models first of all are not you know very easy to derive or even they contain terms that we would have not expected to have we started from a standard conduction equation and now we see effective advection that appears in addition with other terms that look even more complex it turns out that this effective advection is exactly the term that mathematically can model the propagation of a thermal front at the continuum scale that would happen if you had thermal runaway and so of course now we get these models they are produced by essentially our algorithm but how do we check if they are correct so we perform a full numerical verification in this particular case we do a verification case where we have batteries that have all the same heat generation profile and we perform full pore scale simulations this is our benchmark and truth and then we compare with the reduced order models that symbolica provided and if you can see on the left the average temperature as a function of time this is a perfect match between the pore scale the fine scale i fidelity simulation and the continuum scale but our modeling actually our verification step is very stringent and we are not just happy to see a visual coincidence but we actually check the error so we call we say that the model an equivalent model is fully verified only if the error actually is bounded by the theoretical error that is prescribed by the theory so if the error between the two simulations is below this red line then it's a verified case so we can do also more complex cases in this case we have batteries that have different heat generation profiles the battery to the left are already undergoing thermal runaway that then causes the heat propagates to the right and then causes the other side of the battery pack to undergo thermal runaway and again we see that if you if you calculate the error we are all under the theoretical bounds we can even do a full scale thermal runaway analysis where again we have you know half of the side that has already undergone thermal runaway and then the heat propagates to the right and side and we see that we can capture again the behavior but the fundamental difference is that the full high fidelity simulation takes 10 days to run in serial while while this reduced order model with accuracy guaranteed takes less than a day so of course once the models have been validated now we can start making observations or analysis that maybe could guide the design principles and the reason why we can do that is only because we can now rerun all this calculate all these reduced order models while changing parameters of interest I would like to emphasize that the reduced order model that you obtain is dependent on the parameters and the physical description of your system so if we change thermal properties of the battery cell and packing materials or the thermal resistance of the insulation layer or the heating generation profiles of batteries and the state of charge then the equation that you would use might be different but at this point we do not care because we can essentially run and derive these models instantaneously for all these different combinations and the reason why we don't want to do it by hand is because that's how it looks so this is just a conceptualization of how the models can change depending on the parameters and now because of Symbolica we can have essentially all these different models and look at how the different parameters can affect the response at the continuum scale we can even start observing essentially how we recover some models that have been already used in the past so for example for certain ranges of the parameters you actually recover the lamp capacitance models and from this type of analysis we can also identify for what range of parameters we should not be using reduced order models because we cannot guarantee accuracy and this is represented by the range of parameters that correspond to the white area in here and other we can identify ranges of parameters where for example we know we're going to have a negative impact on heat dissipation now the question is this is all good right we have we know that we could use potentially lamp capacitance models that have been around for a long time why then do we need all this well so what we did was actually to check where the parameters are if we consider real batteries and so we did a study where essentially we looked at different types of batteries different types of packing materials and these are represented by the discrete points so what you see is that some of the combinations do fall within the lamp capacitance models but not all of them and so if we want to retain accuracy we really need to make sure that whatever battery or combination of battery and packing material we use we do predictions using the correct model we also emphasize that here we have entire classes of combinations that fall into a region where we actually should not be used or cannot be used reduced order models or equivalent representation so how do so that leads to a number of important consequences so first of all one first observation is that even the veracity of continuum scale or reduced order models can be dependent on state of charge even for the same chemistry so let's say that we know that the reduced order models performs well for a given chemistry in the given state of charge that doesn't mean it will work well for that same chemistry at a different state of charge and we can also perform numerical simulations but i'm going to go a little faster on this so what we would love to do is that if you have a battery pack and specifications and geometry and measure measured heating rates for a given state of charge we would love to get this information from you so that we can construct our heat generation function build the model that we think would accurately represent your specific condition and then try to you know predict what's going to happen and hopefully compare if you have measurements of the temperature distribution in the pack so but what we can do if we cannot use continuum scale models so again it's very important that we have a priority control of the error of our models this is necessary because of moderate reliability and safety and it's never advisable to use models that we know we cannot rely on especially when safety is at risk however we can always go to high fidelity simulations but they are too expensive to simulate the entire system so our solution is to use hybrid models so what hybrid models do differently from high fidelity models that are represented here in figure a or equivalent continuum models that are represented in b is that they combine both representations and so what they do is that they resolve the fine scale only when the continuum scale solution is known to be inaccurate and because of that they still can guarantee predictive accuracy but they're also computationally cheaper than fine scale simulations because most of the domain is still modeled through equivalent representations that this advantage again is that the building the coupling the numerical coupling between fine scale and continuum scale equations is not trivial and requires some specialized mathematical expertise so this is just an example of how a hybrid simulation looks like at the top you see fine scale simulation where we have 10 cells to the left that undergo thermal runaway and then the heat propagates to the right and triggers the other cells in the middle we have the classical reduced order model simulation and upscale simulations for the packing and cell temperature and at the bottom you you see what the hybrid simulation would look like so you have part of the domain that is resolved by using i fidelity simulations and that is coupled with a continuum scale simulation of course the computational cost of these hybrid simulations depend on how big the poor scale domain the fine scale domain is but that kind of analysis can be performed quantitatively and we can really estimate when you start getting significant gains by using hybrid simulations over using full scale full scale high fidelity simulations in this particular case study we looked at a domain with 80 battery 80 cells and 800 cells and what you see is that the breakeven point becomes more and more advantageous for hybrid simulations that the speed up becomes quite significant so we can get a 15 time speed up in a simulation time compared with full scale simulations and of course we are also interested in the error we want to ensure that our hybrid simulations are accurate this is again to the left the simulation of the temperature profile as a function of x the fine scale solution which is the accurate solution is the dots the upscale solution is this dashed line over here blue line and the dark ones are this the hybrid simulation so what you see is that the upscaled or reduced order model simulations really cannot capture the correct temperature at early times but later eventually it can catch up and so what we wanted to do is to essentially ensure that we do not run fine scale simulations unless it is strictly necessary and for this reason we also developed adaptive in space and time hybridization strategies with automatic detection so this is just an example where we use this automatic detection strategy the top is a fine scale simulation and at the bottom is a hybrid simulation with self-detection for an expanding domain so we start from fewer cells being modeled explicitly and then this domain extends because and follows essentially the thermal front we can also do another case which is contraction again the fine scale simulation is at the top the hybrid simulation is at the bottom we start by resolving a lot of cells but then because the thermal gradients decreases then the domain size of the fine scale simulation decreases too and so the errors are always good and this is essentially where we are at and the next steps is to actually go down the scales right now we have been focusing on the cell and the battery module and pack scale but now we want to incorporate physics that go lower to the particle scale in particular we are interested in coupling electrochemical transport and heat transfer to include all these effects up to the battery pack scale and we want to develop rigorous reduced order models for online deployment with these I would like to conclude building accurate and predictive reduced order models of thermal runaway in battery packs is important to ensure safe operations under a variety of abuse conditions and to do so at a lower cost experiments can be very costly developing such models is time consuming complex and error prone yet meeting climate goals does not allow for delay so new theories and technologies developed in the context of formal upscaling symbolic automated computing and self-detecting hybrid strategies can really significantly advance our ability to perform computer aided design to improve safety during regular end-up use conditions and with this I would like to thank you all for your attention I would like again to thank you the group members who provided essentially the content of this presentation and the sponsors Lenya thank you so much for this tour de force of the modeling landscape it was it was amazing to hear and you organize it so well and gave us a real appreciation of the complexity as well as sort of the trade-offs and things to consider starting at a high level just out of curiosity if you were for some of those models when you know sort of the multi-scale and multi-physics models how long does it take to build such a model so that's an excellent question so there are different components right so there is the first part which is actually to derive the equations right and that could be incredibly complicated and sometimes not even feasible if you have a lot of different physical processes a lot of scales and up until very recently this kind of formal derivation was more in the ivory tower toy problems academic examples when then we introduced symbolic automation where now we are not limited by complexity anymore our claim is that now this rigorous model development can actually be done for complex systems we are very aware that practitioners and real world do not have you know the luxury of using complexity as a tuning parameter like we always do in academia and so we wanted to fill this gap and now we feel that with this new technology we can really do that we we can do it really fast we again we are not limited by the complexity of the system by the multi-scale nature of the system practically you can give us as many equations that you want the computer is going to do it for us and you can do it in like one minute it would take before years to do that so that's you know it's been enabled now now then there is the numerical part once you have the model then you have to develop the numerical component that represents the model and again there has been a push to automate also that component in a way that because I said we cannot afford to develop to develop these accurate models and take 10 years before that's what would happen right for co2 sequestration we have this the highly sophisticated codes where most hundreds of scientists have worked on relational labs universities but it takes 10 years to develop them so we believe that symbolic computing can really take a major part in the stage of advancing significantly speed up these processes so that these very sophisticated models can now be developed like in a year or a couple of years maximum from zero from zero all the way to deployment so yeah in general it takes a lot we are trying to you know kind of address that well it's very exciting because now given the urgency and the timing that's necessary to make a significant progress having tools that we can create these models in real time and allow us to actually predict some of these behaviors is incredibly important having that toolkit for us you know it's just incredibly important and it's exciting to see this now so just so that I have a clarity on this so if you were able to develop this full model could you predict behavior at all the scales within that model from all from the macro scale all the way to the micro scale and that's the power of such model is that accurate yes that is correct when we get there because right now for this thermal runaway problem we are focusing on that specific scale right but we intend to kind of actually we're already working on it and intend to kind of propagate down to the to the all the all to the particle level scale yes that's precisely the point if you once you have the model you can essentially predict the behavior of the system at every level uh now naturally validation with experiments will be critical but I think we are getting there in the sense that now recent publications have shown that we can now actually see the dynamics of charging and discharging at the particle level and so you know we you know new technologies and so I think that the experimental imaging is will will and is already kind of there for us to be able to cross validate this model right because there are two parts like the model there is the verification part where you create your synthetic system you presume that the PDEs at the verifying scale are correct and then you essentially derive everything else and that is a self-contained process you can simulate numerically the full scale the high fidelity and then make sure that your continuum scale model represents that but and it's a self-contained conceptual loop right but then of course you got to add the the experiments because you may have in the model you start with some inaccuracies right and so that's where coming out of that universe it's super important and interact with experimentalists because that's when we can verify that the fine scale model we start with is actually accurate because then it's all a cascade of effects that have to be captured yeah so staying in the weeds a little a little longer if you have a complex set of uh fine scale details like if you were talking about a real battery and have a distribution of particle sizes distribution of surface morphologies etc are you able to incorporate that kind of distribution or complexity in such a model yes yes you can and we have done it uh actually for batteries but not for thermal runaway for electrochemical transport and uh i scheme over that part because it becomes very technical but i can give you a flavor of where that information would be so these models these equivalent representations they contain effective parameters um these effective parameters can be calculated by solving a boundary value problem on that rev that i showed you or on the unit cell now the specific information about the topology and morphology comes in that unit cell because now that unit cell you can take from exit scans you can solve your your your boundary value problem that unit cell and calculate your effective parameters and we have done it for uh batteries but for electrochemical transport and we you can then sophisticated the approach and speed up the approach by using even ml and ai which we have done but i have not included the team here yes so the answer yes absolutely you can take realistic images and structures of your battery and incorporate that very explicitly in these equivalent representations wow that's very exciting so in what you're painting is a pathway toward full uh full modeling once you have a adequately characterized and incorporated in your model and uh predicting a range of behaviors so that's actually my next question so we talked to you presented quite extensively on uh say thermal runaway but you know other aspects of batteries are are rapidly developing new chemistries um discovering the impact and importance of stress concentrators all these kinds of things are now being you know evaluated and are stood under importance so if i were to take a leap then in principle those kind of properties can also be predicted once you set up your model for the right application that's correct yes absolutely okay do you have other examples of where you have done this besides thermal runaway yes so we have done it for electrochemical transport so um yeah calculating for example effective conductivity and effective diffusivities of you know battery electrodes using this again i don't want to become too technical but using these essentially upscaling theories and using images and in that case wrapping them around um kind of machine learning to speed up the process and kind of beat uh brugeman kind of relationships um both inaccuracy and breadth of the predictions all over the parameter ranges in terms of porosity interfacial area and all that we have done it also actually for reactive systems where you also have advection so for example if one applicates we have done it in you know as a motivation in the context of you know flow and transport in geologic media for co2 applications but essentially the same equations can be applied for flow batteries and so we are able to use topologies of the structure of the small structure uh to essentially calculate effective dispersion of the material effective you know effective dispersion and and the effective reaction rates right so we can do all that so we have done it and we have published on this but yeah anytime um you have essentially a complex structure the framework is already there and it's only the partial differential equations that change wow that's that's really exciting especially uh if I assume you would run if you have the same model when you have a characterization of a new chemistry yeah you could move pretty fast oh yeah i'm predicting behaviors yes we we only really need images right because essentially we need the image for the poor structure and then what we need uh to actually kind of have the model with the parameters is uh reaction rates so that like it's a model input so kind of characterization of the reaction rates at the fine scale which would come as a boundary conditions in our case but essentially we just need a structure uh effective diffusion in bulk in the bulk prop you know bulk properties but you know without crowding effects um and then reaction rates at the interface so um yeah that kind of characterization and then it it's all kind of I would say for free once you have the framework not really until you don't but yeah and have you applied this to some of the interfacial effects because interface especially with these new battery systems are rather complicated and there are many of them and uh not just at the fine scale but also at the next scale and so on yes this is something so again uh we look at interface I mean you mean grain interface inside yeah grain interface maybe electrodes you know solid electrolyte interface there's you know solid safe batteries there's a lot of yeah so at the moment we have only looked at um solid uh liquid electrolyte solid particle interfaces but again um you know kind of the framework is there we have not applied it to other cases but yes this is something where where you know including grain interface it's something that I was always interested in looking at because again I think that again it's a not it's a very natural process to start looking you know a finer finer finer scale so not done yet but that's where we would like to move oh absolutely and we're finding out how much of the behavior is dictated in fact by those things that are happening at the interface oh absolutely absolutely yeah so let me just take a a step out a little bit which is uh it was fascinating to describe to hear about the multi-scale multi-physics full model and then you're talking about the hybrid model if uh if I would have looked and think about from the standpoint of um savings if you will if you wanted to have a hybrid model uh compared to say a full-scale model um how much faster or easier is it to build such a model so again we are uh I mean you mean numerically um for predictive purposes for predictive purposes yeah so so in our group we are currently working on automating everything okay so because our plan is to kind of have these technologies out of the academia and then eventually being deployed you know outside yeah but to do that you need to retain complexity and we can't wait and so our plan is to automate even the generation of the numerical software so that it comes out as a package that then can be you know kind of uh because the point is that is the development the development takes a lot of time it requires a lot of very kind of specialized expertise a particular combination of expertise and again um and it requires also um knowledge knowledge in the specific field right so you need the knowledge in physics and chemistry if you talk about batteries but then you win or electrochemistry but then if you apply the same tool to I don't know hydrogen storage you would require that specific knowledge right so the automation part is aimed precisely at overcoming the hurdles of having to start over from scratch every time for every application um once that framework is in place again it's very fast in the group we develop the hybrid for batteries in in like less than a year okay but there is a lot of like of course historical knowledge right you know we've been doing this for 10 years right so once you have that knowledge it's kind of relatively fast to implement it numerically but we want to shrink that time and we also want to lower the level of expertise that you need in order to generate your own model for your own specific application yeah fantastic thank you for keeping the line on I know it's weird it's like if somebody walks by actually turns on but if I move a little it doesn't so you know one of the things that is becoming very challenging right now especially as we scale so rapidly is introducing of new technology and validation or confidence in these new chemistries or other things fast enough before you invest these enormous amounts of money in scaling and you know traditionally the the timeline associated with new battery chemistries and scaling is measured in decades so it's you know it's fascinating now that you could have a tool that could help you leapfrog a lot of that learning and you know before big huge amounts of money are spent in in investing new chemistries for batteries or other things so that's that's incredibly exciting because I know that many of the people are trying to struggle with that dilemma of when to pull a trigger and have confidence in in doing that before they make that large of an investment yeah yeah again it's important the model validation yeah so that's uh it's it's it's we we yeah we need to have confidence in the model so that they can you know be reliably used to make those assessments absolutely yeah I I understand that and in many cases it's not always exactly the same when you're talking about a small scale versus a large scale but right now your your modeling gives a gives a pathway to getting greater confidence assuming that you have the same kind of characteristics of your batteries if you scale up and those can actually be validated separately absolutely so that's uh I think that's what makes it so exciting is that you have a way to be comprehensively for taking a lot of the battery performance aspects yeah and also like I think you know if you do like this skill translation accurately you can also identify emergent dynamics so dynamics that again you might not see at the fine scale but then it emerges as a result of coupling processes over longer scales or larger scale so you don't even have to presume how the system is going to behave at the larger scale it just emerges naturally if you know the calculations are performed accurately okay okay fantastic thank you Elenia and at this point I would like to go ahead and bring back Troy to the discussion hi Troy right enjoy the Elenia's talk there I enjoyed yours too thank you Troy and actually I just realized we're both Tritons I didn't I hadn't made that connection I was a few years before you though you're both what Triton you both built your PhD at UC San Diego ah and that's a Triton yeah you know it's not a division one school so you don't see it like but yeah there's a Triton so right all right there we go all right fantastic um so you know one of the things um that we see right now we're uh we're that's a concern as we scale up really is around this amount of batteries and you know what do we do with it and how much residual value is there in those batteries and their second use you know etc and so one of the questions I have which is a little bit um a little bit of a segue but really it's long can you use types of these types of models to predict things like or maybe a residual life and can you do that in a way where there's a small amount of work to tell you how much residual life there is and potentially incorporate that as part of a value statement for second life batteries or second use batteries or etc that's one of the big opportunities that I see coming over the horizon and I'm opening that you know from both and any of your modeling but also Troy kind of understanding what kind of characterization might be necessary to validate that yeah it's interesting uh you know the approach and strategy for recycling versus second use obviously the most direct if you can just take a battery that needed a low impedance and high high discharge capability from an EV and plug that into a low discharge energy storage facility that's great right uh challenge is obvious are when people are using different form factors that doesn't work so you need to have your modules from you know if you have someone huge like a Tesla you can design a system around their their battery modules but um you know there have been forward thinking individuals that said well we can we can use batteries of all types and mix and match them all together yeah that would be a nightmare if there were failures like how do you know even you know which type of cell to start with it would be a nightmare but but yeah it is interesting an interesting challenge to um you know is it is it the impedance that's most important um you know they're each battery because it behaves different based on this design and chemistry um it's going to be a little different so there's not going to be a one solution for all of them but but certainly impedance of the cell is is one critical thing you can look at um your charge efficiency and it to kind of monitor a state of health um but but yeah I don't know Eleni have you done much on modeling of you cells or you kind of think of what what things are new and that's where you're starting everything yeah no no we have not right it's um that's why I put the slide I want you because again all this that we have done uh it's essentially uh you know all kind of theoretical and that and some of it actually was done based on uh you know published you know data but that's why I think the collaboration with experimentalists is so critical because I think that we could actually help answer the question and I think that there is an additional component into you know the complexity of this is that in order to be able to predict you know kind of battery life numerically you have to run a simulation that simulate battery lives right so it goes back to that to that separation of temporal scales where now we should be able to build models that actually can be run efficiently to simulate years of operation right but again I think the framework is there but in this and we can certainly go off a tangent and do it but I think the integration with experimentalists and understanding like where they are coming from what is it that they can or cannot measure right because uh one thing that I found in the past working with experimentalists in other fields is that with modeling I mean we have it easy to some extent you know we have our you know uh whatever virtual environment and we assume we know what we want to know and we then do predictions and make statements but oftentimes we forget that not everything that we want to know it can actually be measured and so you know there has to be a and that is very important for model validation which I think is again super critical to build reliability in the model so that when a model is deployed to provide advice or guidance can be relied upon and so so I think yes it's possible you know there certainly will be technical hurdles on how to simulate thousands of cycles passed without having to spend years in simulating time or you know whatever months but it can be done but it's important to then iterate back with the experimentalist to see like what is it that they can test so that we can build confidence in these models before starting making assessments or you know provide guidance. Another thing on that Jim you you talked about what different things can cause their runaway well you imagine we're talking about handling new cells and new battery modules and the assembly can cause some headaches and cause some issues now imagine you've got a cell that's gone through you know 10 years of swelling it's got more stress in it if you have to do a second use at a cell level now you're going to be taking out cells that don't have the same rigid rigidity they have different degradation and so whether you're doing a module or whatever level you got to consider those things my my phd was on the failure propensity of spent nuclear fuel things that you were trying to handle nuclear fuel rods after being in dry in wet pools for for many years and then dry storage before going into you know a permanent repository they're degraded to no end and this huge challenge trying to work you know work with the alloys that have had irradiation and creep and and so the same thing with batteries just like now you've got a whole new system whether you're modeling it or designing your your structures to deal with degraded and you know behaviors that we don't really know because we're still in the front end of this process for us it's just a condition yeah well it's very exciting now because many of these applications now are actually are actually profiling the battery used over the course of their lifetime so now they have all this data and uh and they're potentially now in principle have given you a path to start looking and you know understanding the impact of these various things and if you're able to simulate it over lifetime like you're describing and know the use profile discharge etc charge and you know it would be amazing to be able to have models could actually predict okay under these kind of conditions we expect you know this to be a this is your life value of this sort of or many lifetime of this with this type of application sort of a report card if you will that would come with it and it could be at the pack level but you know strongly I could see that being tremendously useful yeah and we do this for other applications right so for example for co2 sequestration I mean you look at uh you know centuries of of you know in the future you want to make sure that it doesn't leak out right so the technologies are there to actually do that and I think that we just have to take advantage and uh of you know our know-how from other fields uh and then essentially apply to you know this this this specific domain knowledge and it's also one of the things we battle to be honest is it's not a big academic setting where we're just doing research on batteries there's also the commercial competition and so a lot of these data that you're talking about are there and certain manufacturers whether the oems like the you know gm and ford and uh crisis of these they have data then you have the lg and panasonic they have data they they know their own things they aren't rather necessarily willing to share what is that key piece that I've learned from my research with everyone else because if their battery can behave safe more safely or or longer life that gives them a competitive advantage and it's really interesting to watch when you go to the conferences and discussions you you kind of move that direction people are very secretive and that's kind of our ip and then when you have these penultimate events like the the sony we call back in the the you know the late 2000s and i think recently on something like the the big gm lg recalls then it's more of a community that people feel hey we're being attacked for the integrity of our batteries as a whole because the top-tier manufacturers are failing and so now we're losing the public you know uh confidence and so this whole technology may not succeed if we don't work together so then you kind of shift back to sharing and kind of doing the right thing but then you kind of forget about those events and pretty soon it's it's kind of this competition ip so it's kind of interesting to watch those cycles yeah i yes absolutely and certainly data right now is one of the things that people hold very close to their uh to their vest and uh you know we certainly see that also but i think that's where university actually can play a role of just disclosing everything that we know because for the you know for putting the knowledge out there so that it can be used uh you know to improve uh society um so yeah i think uh yeah can work yeah absolutely and i certainly think here uh you know as part of uh the new stanford doors go sustainability that is part of our mission really and so it's very exciting to see that um a topic around you know um climate uh meaning regional climate differences so if you do some sort of characterization of battery performance and you do it under a certain set of conditions but then the climate varies dramatically uh like the climate in india is going to be very different from the climate in paloalco okay the climate in uh in saudi arabia or any other regions or if you go way up north and so my question is uh along the ability uh of these models to predict those kind of climate effects on the performance of properties of the battery do you feel like at this point that you could actually you know plug in and says well if i had this type of humidity humidity level or just kind of temperature variation over the course of a day or you know a year or something yeah i mean again we are not there yet but theoretically yes because these are boundary conditions for us so once the infrastructure is all there again we have it easy to some extent like once we build a complete infrastructure then of course you could start accounting for all these kind of effects and for us those are essentially boundary conditions so we just impose those and um and run the forward model um yeah but we are not there yet so it's you know we have boundary conditions clearly but we cannot account for humidity right it's i our boundary condition is just to external temperature and that's but that's already that already can be included yeah and these outliers jim uh jimmy it's it's it is really important um if for those that have owned evs they they frequently can turn themselves on and use their own air conditioning and things like that to keep the battery at a certain temperature range but what happens if you drive into the phoenix airport and you're down to 10 state of charge and you go on a business trip for a month um you know is there enough remaining capacity to actually continue to to keep that in the condition you need or is it going to see these spikes and if they have these spikes how are you how are you dealing with it um you know on the management of the battery moving forward yeah we we've we receive these kind of requests a lot which is you know um we understand and read about the better performance in say Palo Alto okay you know but we're we don't have the same climate as a Mediterranean climate in Palo Alto around we have this and that and stuff you know we're interested in knowing how how it would actually perform in this in this climate and stuff and in a lot of times i know a lot of people are running a mountain of experiments characterizing them under different use conditions different climate so i would imagine if there was such a model which they could could predict it would be tremendously valuable as this spreads around the whole around the whole planet and so i have a changing so that becomes another thing right it's dynamic now right and so it's going to happen in 10 years so i have a question from the audience um so let me just read it um would these models be leveraged to determine gas compositions expected during thermal runaway knowing gases that are that are present is critical for fire safety design and the only way to get that data now is to literally to do thermal runaway and measure so so we don't have the capability yet but that's why i included that comment on uh framework flexibility and generally generalizability sorry i didn't get that right but anyway uh so the idea is that as we have the infrastructure this this framework then we can start adding physics uh so at the moment we have demonstrated the capability of the framework you know essentially on a basic benchmark which is you know conduction and the heat generation and heat sinks associated with for example cooling pipes but then ultimately the idea is to add layers of physics um that can then be modeled uh to include additional effects we don't have this capability yet but the idea is to increase the complexity to get all the physics that we need uh so that they can hopefully you know match the experiments that's a choice running yeah and one thing that's really helpful there is of course you can measure the gas volume and composition um but also nasa designed a fractional calorimeter which we use also which allows you to differentiate between the positive material ejecta negative material ejecta the gas um and what's remaining in the can so you're able to kind of really drill down and get data for those models um and that's that's going to help us as they get developed moving forward well that's that's amazing so you could when you say physics then do you also refer to chemistry oh yeah oh absolutely okay that's i think that you know our our current work actually zishin um the student is working exactly on coupling the electrochemistry with the thermal effects because the idea is that this heat generation curve is actually driven by the chemistry and so the idea is that then by including you know the electrochemistry part uh then you know that becomes all fully coupled and integrated because clearly the temperature rise you know it's what leads to this more exothermic reaction that then gets sped up and then produce a further uh temperature rise right and so that coupling I think is the one that we want to capture first uh there are so many things that we need to do but I think that this is the one where we are investing now most of our effort coupling the electrochemist electrochemical transport with uh thermal amazing so uh thank you on behalf of our community here in storage acts and people around the world that listen to us thank you very much for your contributions and efforts in this space it's great to and for participating in our storage x symposia it's great to see the progress and the learning and the sharing as you as you both are doing uh for our community so we have our upcoming events uh we have a next week we have our second international symposia for the summer frisk graves and column vessels uh these are two startups that are now within this ecosystem uh so come in here about that and we also have our storage x talk uh tuesday next tuesday uh by john hollow back here at stanford so please join us for those two events and we look forward to continue the engagement and so again elenia and troi thank you very much and we look forward to hearing from you all right bye bye everyone