 Good morning from Stanford University. My name is Welchu. I'm the director of StorageX Initiative and a professor in material science and engineering. I would like to welcome everyone to the final symposium for the winter quarter. Today, we're discussing a very exciting topic on battery management, prediction, and diagnostics. So let me give a bit of context on this topic. If you have seen a battery gauge on your computer, on your smartphone, or for your electric vehicle, you have benefited from the expertise of the two speakers today. In addition to being able to determine the state of the battery and the health of battery, increasingly it has become important to also accelerate the pace of research and development for battery technologies, especially for lithium ion ion. One of the reasons why it is so widely used is because of its lifetime. But this long lifetime also makes it very difficult to predict the performance of the battery and has been a bottleneck for the research and development process. So the field of battery informatics has been growing over the past decade, where large data methods are being used to accelerate this process. And I am delighted to be hosting two of my colleagues who are world experts in this area in battery management and prediction and diagnostic to talk about various aspects of this very important problem and arguably one of the underpinning technology needed to speed up the energy transition. So we have today with us Professor Simona Onori from Stanford University, my good neighbor right here. And then also Professor Dirk Uwe Zauer, who is a chair professor at the RWTH University of Aachen in Germany. So to get us started, let me also introduce the director of the Pre-Core Institute, Professor Itui, and he will introduce Simona. Itui? Well, thank you, Will. It's great honor to introduce Simona, our young, but really a young star colleague that's why he has Stanford University. Simona joined in Stanford several years ago in the energy resources and engineering department soon. Everybody is going to see we are going to have a new school on climate sustainability. I believe Simona will be moving into the new department called Energy Science Engineering. Simona has done really nice work on battery management. I actually benefiting from interacting with her quite a bit. She has won many awards. I will just mention the NSF Career Award. I look forward to many awards. She's going to win down the road. And Simona, take it from here. Thank you so much for the nice introduction. Thank you. Do I come through right? Can you guys hear me? Yes, I can hear you. So good morning, everyone. And it's a real pleasure and honor to be here today in front of this very educated audience to talk about battery management system. And I'm also very excited to talk today with Professor Sauer from Germany, who has been a leading expert in the field. And as Will has mentioned initially, we are going to talk about battery management system. This is, I would say, a novel topic, a new topic within this platform, the Stora Jax International Symposium. So thank you for the opportunity to share with you, with all of you, some of the work we are doing in my lab. At Stanford, I lead the Stanford Energy Control Lab. And I have the pleasure to work with talented students, postdocs, undergraduate students, and the visiting scholars. And broadly speaking, what we do here is to develop algorithms, so software codes, and to make energy systems more efficient and last longer. So we really care about trying to extract the most of the performance out of the system once they are deployed in given applications. The majority of the work we are doing today is centered on lithium-ion batteries. Although we also look at technologies related to after-treatment systems, like three-way catalysts and gasoline particular filters, and through NSF support. So we are, broadly speaking, a dynamic system control optimization group. And I would like to show you the type of things we look at besides and outside the BMS. So hybrid electric vehicles has been the focus of our research for a bit now. And in hybrid vehicles, you have two sources of energy power. And one of the main challenge there is to how to decide which actuator to use at any given time. And over the past years, we look at formulating energy management strategies for this problem. And what we're doing today, on the other hand, is to kind of reframe all that problem in terms of exergy. Exergy is the reversibility that you get in the powertrain and in the components. And we're looking at modeling the exergy dynamics within the powertrain and also deploying some control solutions. Electric vehicles are being analyzed and studied. And within this type of vehicles of systems, we look at driving cycle analysis, understand driving cycle dynamics, and to bring that knowledge back to the operation of the battery and to understand how this information can benefit in terms of design a battery management system. We look at different opportunities to hybridize or electrify the system, the vehicle. And so we let your battery sometimes might not be the best and only solution to use. For example, in semi-electric, electrified semi, or even military vehicles. And so we look at opportunities in that space where we combine, for example, supercapacitors, solitum and battery technology, and also create optimization routines to optimize their size. Battery packs is the most important components in electric vehicles. So we look at the BMS battery pack systems in terms of designing control estimation and optimization strategies. And we run high fidelity models to understand the dynamics of a battery pack. We use also, we are looking at understanding field data from BMS and using machine learning algorithms to interpret and filter and process a huge amount of data that we get from the vehicles. Cells are the main component of a battery pack. So we work a lot around understanding and modeling the battery cell. And we develop physics-based models, implementing code physics-based models of different flavors for battery cells in terms of electrochemical dynamics, thermal, and aging. And we use those models along with the machine learning routines to assess the health of the battery, understand the amount of charge in the battery cells, and also predict remaining useful life. We just started a new collaboration with a new sponsor in the Michigan area on battery cell manufacturing. That's a new area for us. So we're very excited to contribute with our work to the making of new electrochemical cells. So all of this work is done tackling and using fresh cells or new devices. But there is an impending problem today which is reusing and recycling on lithium-ion batteries. So we also look at this challenge in terms of understanding feasibility of used batteries coming from retired battery pack from EV into grid storage as a grid storage solution for smart grid application, for example. And so one of the challenges is that when we take those batteries out of electric vehicles, there is really not much trust in what the BMS predicts in terms of battery life. So the main question is, what's the health of those batteries? Are they good to be used in a grid storage application? And if so, which grid storage application should we use in? And how long do we expect the battery to last? That's a huge investment. So we need to be able to give those quantitative answers to those questions. So my talk is going to center on a battery cell and battery pack today. I want to just quickly go over some high-level information here. In the US, we feel very strong the support of our government towards electrification of the transportation sector and the power sector. This is important to reduce greenhouse gases get away from petroleum and therefore be diverse, especially and mostly also to improve national security. And we read this past two weeks have been very difficult. There is a war happening in Ukraine. And we are feeling the effect of this war in our daily life. Electric vehicles can support local communities. So low-income and disadvantaged community can benefit from, for example, electrified buses. And one thing that is also going to happen is that the individual ownership of EVs is going to be coming possible for even more people. This is a picture it took almost in 2020. That was September 2020. And I'm sure people still remember that week. It was a horrible week. The sky was red. And this was taken at 2 p.m. on I-202 AD when I was driving from San Francisco to Palo Alto. And the air quality was very unhealthy. And that was a very scary situation, I would say. And that was the effect of a big heat wave that we had in Northern California the month prior. And that has revealed that our electric grid was not capable to, or not even flexible, to provide the demanded energy. So energy storage here can be used to create, to alleviate those situations that are happening more and more often in the past years. The huge commitment towards electric vehicles is not just from the government, our government, also government around the world, but also from automakers. Automakers are making huge investment into this technology. And if you have watched the 2022 Super Bowl last month, you have noticed that there were more and more electric cars features in the commercials. And for the first time, there was also a Powerwall charger feature. So they're really getting our house also from other channels. So that's a very, indeed a very good news. Lithium ion batteries are the key technology to unleash the market, the mass market potential of many, many systems, not only electric vehicles. And we have seen also with consumer electronics and electric vacuum, but also good storage and beyond that. Heavy duty vehicles, we have electric boats, and also unmanned vehicles. So it's happening. It's around us. It affects our daily life. And every minute, I would say, our daily life. Now, if we were talking maybe like a few months ago, we were praising the declining price of this technology. And we were also commenting on the fact that this trend was a trend that we were used to see also when it comes to solar technology, PVs. Now, unfortunately, and so the price actually of lithium ion batteries technologies has declined almost 90% of the past decades. Now, what we have seen recently, what we read on every day is the supply chain crisis that has led the cost of metals using the cut material of this technology going up to the roof. So the price of cobalt has doubled. Nickel is 15% more expensive than the beginning of the years. And I'm sure today there is going to be some other news already to this increase of prices. So what's happening now? So some of the forecasts says that this year we're going to have a reverse trend in terms of battery price. And we'll see how this is going to play out in terms of EV sales. On the other hand, there are also new opportunities that are coming up of this situation, one of which, for example, is a new market of reuse EV. So we're going to have, we're going to see more and more used EV sold dealers. And so that will increase the EV ownership. So last year around, I think it was June of last year, the DOE issued the National Blueprint for lithium ion batteries. And this is a very nice document that you can download from the energy.com website that was put together by the Federal Consortium for Advanced Batteries. And what you see there, what you read there is a very nice list of goals. There are five goals that are very well stated to improve the domestic manufacturing and production of the materials for batteries, manufacturing of electrocells and packs, and the ideas to try to really be independent from other countries, especially countries where there is some political instabilities. And when I was reading this document, what really got my attention was the fact that there was a missing piece, which I believe should be added there. And that is the battery management system design. BMS design is such an important component of this technology that should be recognized and acknowledged as such. And there is a lot of engineering around this component and that also requires people to be educated in this field. And I hope to see this goal, this new goal in the next in the future. So different from what E and my colleagues, Ian, will do on a daily, which is designing better batteries by discovering new materials, by optimizing, for example, cathode materials and so forth. What we do in my lab is to use control, use modeling to maximize performance of the battery. So how can we use the batteries better and to make them last longer? And so what we need to do is to recognize the fact that around the battery cell and battery pack, now there is a whole new framework on new infrastructure, which is a set of sensors where in packaging and communications and that needs to be incorporated into the study of the design of BMS. Quickly speaking, I want to just mention that when we talk about lithium-ion batteries, we talk about family of chemistry that mainly different from the materials using the positive electrodes. And depending on the application you're using this technology for, you might want to select one technology versus the other. Although there are also other implications like cost and supply chain crisis that needs to be accounted for. But one thing that I wanted to have your attention on is the fact that not only different chemistry is, for example, in the form of NCA or MC and LFP, they have different specific energy, specific power content when you excite them at different C rates. So how can we do current and critical operation? But one distinct difference among these chemistries is their open circuit voltage. So it's the potential that is measured out of the battery terminals when the battery is pressed as a function of the amount of charged stored in the electrodes. One thing that I want you to notice is that the LFP, which is the green curve here, has a pretty flat behavior as opposed to this almost linear, nicely linear behavior that, on the other hand, you would see in the NMC and NCA. So that's a very important thing to notice when you go ahead and create your BMS, because that will lead to, I wanted to warn you, these two observability issues and the ability to estimate accurately state of charge. So one of the things we control, a battery system engineers who look at is also look at these features and these signatures to understand our ability to do a good job in estimating battery metrics. So I would say that if you can, this audience have been exposed to lots of work, great research work ready to battery cells. But I wanted to give you a breakout of what the battery packs look like. So we start with cells. Cells can be coming in the form of cylindrical or pouch or even prismatic. And they're made of anode, cathode, and separator. And the cells then are assembled together either in parallel or serious configuration to create a module. And here is, I don't know if you can see this, but this is an example of a module. This is coming from an electric vehicles, whereas this is the 1865 cylindrical cells. There is a bunch of cells in these modules. So modules are put together to create a battery pack. And the battery pack needs to be equipped also with the BMS, which is this red box here. And then you can put more battery packs together to create a battery system, for example, for electric bus or grid storage solution. One thing to notice here is you also probably know is that there is no standards in battery pack design. There is no standard in cell selection. And there is no standard in module design. We each battery pack use different module configurations. And that makes it more difficult for using and recycling activities later on. So this is probably something that we can work on all together to create some standards to improve in the name's stronger, more robust circular economy. So let me go into BMS. What does the BMS do, what it is, and what are the information that we give to the BMS to operate properly? So let me tell you that the BMS can only rely on very few measurements, which are current voltage and temperature. And temperature is provided from discrete, very optimized temperature sensors, which are all over the battery pack. Current and current voltage sensor are more local. But for example, in a series of configurations, you can have the voltage sensor in each cell. In a parallel configuration of cell, you have one voltage sensor to measure the voltage of the parallel. So there are this type of measurements that we can use to create functions which should make the battery pack safe by avoiding the battery to be overcharged, over discharged, overheated. At the same time, while we prevent through the safety operating area, we prevent the battery to operate in unsafe regions, we want to get the most of the performance of this battery. So that basically goes into controlling the set of charge windows of operation, what is controlling the charge current, making sure the balance is done properly. And then diagnostic and then health monitoring is very important to guarantee safety as well. So AC made instead of health, whatever state of health definition we want to give it to. And I'll cover that later. And then also communication. The BMS needs to communicate with the AC chargers, for example, it needs to communicate with the driver and other ECUs within the vehicles. And again, we are very demanding when it comes to BMS functions, although we give it to very, very little. And that's why, as I said earlier, there is a lot of engineering around it. So two main problems we have in electric vehicles is that when you drive it, you need to know, you need to be aware of how many months you can still drive. And that does not come from direct sensor information. There is no sensor that tells you exactly how much charge is left. You need to estimate it. And so the state of charge estimation problem is about giving the best information possible about how many months you can drive. At the same time, another problem that you need to face now with electric vehicles is the ability to do diagnostics and estimate health. And so health is related to the amount of usable life that the battery has left. And there are tricks that are being implemented in today's batteries to make sure that this health is improved. And so battery is oversized so you can rely on the extra range of energy that you don't use at the beginning of life to release it later through some update of the software. But that will last just as much at some point you will see the critical performance of your battery. At that point, you might need to sell or return the vehicle. But this is the oversized of the battery is, for example, designed and engineered to make the driver have a good experience driving the vehicle for 8 to 10 years. Now, those two quantities, as I said, are non-measurable. And so on the other hand, they're very, very important. So what we do is to create models and equip them with control and monitor algorithms to address these two very important questions. We have a holistic approach when it comes to BMS design. I'm going to go quickly because I don't want to take too much of my time. But we really take this from start to finish. And so we do experimental testing. We collect experimental data of different cells. We create a model and we have quite wide spectrum models we created in our lab. We have many libraries spanning from electrochemical to equivalent circuit models. And we tackle identification challenges and to parametrize those models accurately, which we then use for model-based estimation, especially related to pseudo-charge and pseudo-health. There is some side work we do for temperature estimation as well. Now, this is done at the cell level. And then because, as I said, the battery pack is the interconnection of cells in series in parallel, we look at creating high fidelity models for battery pack simulations where we interconnect, basically, we integrate electrochemical thermal aging dynamics in a proper fashion. And then we use this framework, this model-y framework, to develop control strategies, which I'm going to talk later. And then lastly, we are very much interested in the feasibility of our algorithms for real-time application, for real-time deployment. So we do battery in the loop. We are doing the loop evaluation. So our work is really about using tools at the intersection of electrochemistry, model-based control, the driven method, and experiments. And then I would like to, if time allows, I would like to go over three main applications that I picked for today's presentation. And one is the electrochemical, pseudo-charge, pseudo-health observer design. We also look at some machine learning-based pseudo-health estimation, and then optimization for laugh extension, but laugh extension done at the battery pack level. So there are many models available out there. When it comes to design a BMS, the very first question is to ask yourself, which model should I use? Models can vary in complexity, and of course, also in accuracy, going from atomistic down to empirical models. In today's BMS, empirical models are the tools used. And whereas for design, this model here, and this class of models are also, on the other hand, they've been used. And I have to say, this sort of the art is also in today's model is to use machine learning information to equip those models. Now, the difference, the main difference between the ECM and the equivalent circuit models, electrochemical models, are in the fact that these are very easy to run in real time in the sense that from a computational standpoint, you have two or three ODE's, ordinary differential equations, that you have to solve. The main disadvantage of these equivalent circuit models is that they don't embed information about health or aging or any physics whatsoever. And they also require quite extensive calibration effort that requires a lot of experiments. You cannot extrapolate the behavior of ECM outside the condition over which the model has been calibrated. Electrochemical models, on the other hand, retain the physics of the battery. And so as such, they require a much smaller set of calibration experimental data to be calibrated. Their behavior is physically retained. They retain the physics, and as such, they can allow some sort of, to some extent, some extrapolation. So what we're doing in our lab is to kind of challenge the status quo and we abandoned the use of equivalent circuit model for state of charge and state of health real time estimation. And we embraced the richness of electrochemical models for the development of the subserver. So there are a few electrochemical models out there. The most commonly used is the Doyle, Fuller, and Neumann model. This is the more complete electrochemical models that relies on mass conservation in the solid electrolyte phase and charge conservation in the electrolyte and the solid phase. There is an assumption made there which is that the particles are spherical, which is kind of a good approximation for certain chemistry like NMC, but it's not a good approximation for LFP, which on the other hand, they have a more squarish type of particle shape. We can simplify the complexity of the DFN model by neglecting the charge conservation in the solid phase. And within this framework, we assume that the particle is spherical. There's only one particle in each electrode. And in the electrolyte, as it's on dynamics, so we model the charge conservation in the electrolyte phase. Going down to the spectrum here, the single particle model makes the assumption that the electrolyte dynamics do not change. And so they are basically all we care about here is the mass conservation in the solid phase. This is a good approximation if your system is operating on low C-rate. And so SPM and ESPM, they perform very, very similarly, basically in the same manner when the C-rate is below, for example, 3C or so. And it is a good approximation to this battery. Battery packs are getting bigger and bigger. So for grid storage, I would say that SPM is perfectly valid electrochemical models to use if you want to do some in-depth study. So I want to say that the key to apply our observer design is to translate the knowledge coming from the governing equation, the mathematical expression of the governing equation, into the state space model. So ESPM or SPM or DFN comes in the form of partial differential equations. And when it comes to create control algorithms, what in reality you want or you would like to have is a system that is expressed in the form of ordinary differential equations. And so by doing discretization in the solid particles in the electrolyte direction, we can now translate PD model into OD model. And the input, so the control input of this model is current, and the voltage is the output that we measure. So we have access to voltage sensors. And the output, the cell voltage is a function, is expressed as a function of the open circuit potential and the electrolyte resistance and the battery current as well. Now an important step is parameter identifiability. What my student Anirudh did as part of his dissertation is to develop some new routines to tackle the identifiability problem. And those models can come with a big set of parameters, like large set, like 20 or even higher set of parameters. And sometimes it's difficult to really get the handle of what's truth in the value, medical values, those parameters which we try to identify. If you try to identify all of them at once. So what we did is to develop this routine that relies on local sensitivity correlation analysis and equipped with additional cost function. Additional cost functions, what they do, just to give you an idea, to import new constraints. And those additional cost functions comes from adding virtual measurements in the multi-objective optimization problem that is formulated to carry out the parameter identifiability. Details are in this paper if you're interested. And so the identification results and validation results show a good RMS performance. And so once you have a good model in your hand, you can go ahead and try to apply some of the control algorithm to do the estimation you need to do. Now there are a few things that I wanted to work through to make you understand why you really need to go into a little chemical models and why we cannot do something more like a simplistic. So the status quo is a current counting method. So it's a very simple, simplistic model. And what it does is basically take the current that we measure from sensors and integrate the current in this fashion and to get the status charge. So this charges this a dimensional quantity, which is normalized with respect to the maximum charge that the battery can store. There are three main problems with this model. First of all, you need to know the initial condition. Any time you do an integration, you implement an integration operation, you need to know where to start from. Otherwise, your final result is going to be shifted or biased. And sometimes this is not an information that we have. And going back to what I was mentioning earlier on the LFP battery that has a very flat OCB. So state of charge and OCB, they are correlated. And so one way to look at the initial state of charge is by inverting OCB map. And that can be done only when the battery has been addressed for some period of time. Now, if you have an LFP and you have this flatness, you'll see that inverting the OCB to get information about SOC will give you a wide range of values. So even if the battery has been addressed for a period of time, you might not be able to get the initial condition accurately. And so this method might lead to this method based on quantum counting might lead to big errors. Another problem is the fact that you are integrating a signal current that is affected by noise. And so eventually ultimately, you will have some drift there. And lastly, you are dividing this by a quantity that changes over time. Q is this, for most, is the feature that indicates that the health is not the only one. But it does decreases as the battery ages. And so if you don't know this quantity accurately, you are in trouble. So you will get a very off-estimation or calculation of the state of charge. And so there is a very hard, dark solution that is implemented in use today on vehicles or in BMS, which is people use a second order if you want a second model, properly parametrized. And they use the quantum counting method in the lookup tables to have information about capacity and to basically return the estimation of state of charge and state of health. What you want to do on the other hand is to really take all of that out and reformulate and basically bring new ideas into this problem. And SPM, so single particle models equipped new control solutions, is the solution we are proposing. So and that's what we are going to quickly, quickly see. Because this audience probably has not been exposed to estimation and observer design problems, I wanted to give you very quick points and knowledge about the estimation problems. So whenever you have a dynamical system, for example, that it can be anything else, the dynamical changes upon being excited from an input and relying on certain measurements. So if you have a dynamical system and you know its inputs and its outputs and you want to access the internal states, so what you're trying to do is to, in those states are not measurable, what you're trying to do is to ask yourself, can I build or design a model-based observer? Can I have access to the internal states from only information of what I see outside the system itself? And so a model-based observer is nothing but a mathematical copy of the system. And so because of this is a mathematical copy of the system, you understand why we are making so much, we're putting so much effort in the fine and selecting the right models. And that is for the ultimate goal of design the observer. And so you can eventually be able to estimate those internal states by building this model-based observer. Now, there is some fundamental question you have to ask yourself before you can build the model-based observer. But the way a model-based observer, a close-loop model-based observer works is by creating a copy of the mathematical dynamics of the system and injecting the input, the current and the voltage of the system, the measure voltage into the observer and extracted an estimated voltage response and estimated state vector. Now, this mechanism works. So the estimate x hat converges eventually to the real state that you have in the system that you cannot measure because we use a feedback mechanism that relies on comparing the real measurement y voltage with the estimated one. So this error is fed back to the observer and it gets corrected through the gains used in the observer to lead to the real estimates. Now, the very first question that you need to ask yourself before you go ahead and build this observer is can I even build the observer? Is the system observable from the input output? Can I get access to the internal states by only those two measurements? So there are these questions is answered by using nonlinear observability tools. And I'm not gonna go into the theory but I wanted to just explain what the difficulty is here. So you have a battery to two electrodes and those two electrodes can be seen on top of those two buckets, which are, you know, can be filled and empty as the battery gets charged and discharged. And what we do, we measure the difference, okay? So the voltage measurement that we make at the terminal of the battery is the difference of the open circuit potential of the cathode minus the open circuit potential at the anode. And so you understand that there might be for one measurement, there might be many combinations of UP and UN that are gonna be matching that measurement itself. And so that's one big proof, you know, one of the main problems. The other problem comes from the fact that the open circuit potential, for example, in this case of the graphite is pretty flat. And so there is some uncertainty there in understanding what the real value of UN is if you know the data N. So that's the observability issue that we have to deal with every day when we build the observer for batteries. And so we can estimate the potential difference between two electrodes, but we cannot estimate individual electrode potentials nor the lithium ion concentration. We've done some studies and we have shown that observability of single particle models is a function that depends on the current amplitude and the current dynamics. And that's because the current gets into the lead bracket formulation of observability metrics. But we also have shown that observability depends on the battery chemistry, okay? When it comes to use the OCB to get information about the concentration. So having said that, I wanted to make the point that a closed loop observer as a generally traditionally design cannot be used for the system for exactly that problem that I showed in the previous slide. So we had to create a new framework, a new idea, and that will tackle that challenge. And so the idea here is to use an electro-based interconnected observer. And what this is about, just to give you the main idea behind this new design is that we create two parallel observer that work separately on each electrode. So we have a cathode observer that we build that is based on a closed loop cathode and open loop anode estimation. And at the same time in parallel, we have an anode observer. What I wanted to also stress is that whenever you have an open loop observer, the estimation from the open loop observer is as accurate as initial condition is. So in order for us to really rely on this open loop estimator, we need to make sure that the initial condition, for example, of the anode concentration used here in the cathode observer is the correct one. And so, and that is fed back from the anode observer. And the same for the open loop cathode, the initial condition of the concentration of lithium using this observer comes from the cathode observer. In this way, the two observers at each other and the lead to conversions of the estimated states with respect to the real estate. Now, one other point that is very important to keep in mind is that whenever you look at the observability opportunities within your system, you might have the successful answer that, yes, your system is observable when you look at the observability metrics, which is full rank. On the other hand, you might be dealing with a poor condition number. So condition number is a matrix, a feature that tells you how poor the estimates can be from whenever you're trying to design an observer. And so one thing that I wanted to stress here is that with this interconnected electric-based observer, we improve the condition number by an order of magnitude of five with respect to a traditional observer for both electrons. So here are some simulations that shows that even if you initialize your concentration and the bulk concentration of lithium in the positive and electrode incorrectly, you converge. So behind this framework, there is some rigorous Lyapunov-based proofs that guarantee the conversions of the estimated states to the real estate. So some takeaways from this study are the following. So first of all, non-linearities matter. We need to acknowledge and keep those linearities when it comes to build and design the observer. And that's because the observability is really affected by it. Battery chemistry, type of chemistry and the technology you use can affect observability properties. And so that needs to be taken into account in your design. And the electro-based structure overcome observability issues and behind this design, there is a rigorous proof of convergence that guarantee that the fractured estimates converge to the real internal states. I would like to check with Yi about time. I have some more slides to show. Simone, you probably will need to wrap it up soon, yeah. Absolutely, absolutely. Let me just go quickly over the extension of this observer to estimate health. The battery, as the battery ages, the parameters of the model use change as well. So if you're trying to use that estimator that I show for an age battery, what happens is that there is no convergence. So you will see that the concentrations of lithium, in this case, surface concentration does something completely different from the real one. And so what you need to do is to really make some changes to that framework and equip with some adaptation mechanism. You need to be able to adapt some key parameters that change with aging that will infer that will basically advise your system about what's going on. So adaptation laws are used in this framework to basically inform the single particle models in the observer design that aging is happening and so there is some update to be done. Let me tell you that as far as aging and I wanna just use this slide as a final point for my talk. So we're dealing with fresh cell in this case. And so what the community agrees with is that solid electrolyte interface is the main aging mechanism that happens at the anode that the negative electrode as the battery is being used in whatever application. Now this growth in the solid electrolyte interface layer is observed through some capacity fade and power fade. Now, and what that means is that in our model, the parameters change will change with age and the one because that's what needs to be updated. And as this SELA grows, what you will see is a decreasing cell capacity. So cell capacity decreases, but that's not the only things that happens. And so what we look at, what do we incorporate in our model is also an update of electrochemical parameters like porosity, porosity decreases. That lead to power fade. The solid phase diffusivity in the anode also decreases. Transport parameters in the electrolyte also undergo some decreasing trend. So ion conductivity and diffusivity in the electrolyte. And then at the same time, the ion conductivity in the SCI decreases as the lithium go basically as to go through the SCI layers. And also the electro stoichiometry values will undergo some changes. So all of those changes in the electrochemical parameters lead to an increase of electrolyte resistance and also increase of SCI layer resistance and which lead to power fade as well. So one things that we did in our work and what the student did, was to create a new functional relationship to connect to lead capacity fade to power with power fades through these electrochemical parameter changes. So the observer I showed earlier is basically improved and is reformulated by adding now some adaptation mechanism in the form of anode diffusor coefficient in the anode observer and the SCI layer ion conductivity in the cathode observer. And so not only now we can estimate lithium ion concentration in both electrodes and therefore also inferential capacity and the power fade and we can also track changes in those parameters as well. I would like to conclude with this simulation results quickly that shows that even if the observer is wrong initialized there is convergence of the estimated capacity within a very small percentage within the real one. And we did some robustness validation. We inject noise in the current and voltage and test the observer. One way to think of our observer is a low pass filter. So naturally filters out noise and also it's shown to be robust against constant biases in the voltage and current. And lastly we did the bothering the loop implementation. We connect the ultimate we want to really test this algorithm in the real time. We test the battery, we connect the battery to a DC load and through communication, we basically, we flash the algorithm in the macro auto box in the space, ECU, electronic units and this is a real time embedded system. And then we look at and we monitor the state of health of the system itself. That's the layout in our battery lab. And these are simulation results from our battery in the loop simulations and the battery capacity converges and is within a 2% error from the real value. And with that I wanted to just really finish this talk with some takeaways. We, in this work we address quite deeply the problem of observability and the feasibility of classical observer design when it comes to estimating a pseudo charge instead of health. We believe that electrochemistry equipped with control theory and rely on experimental data constitute the state of the art of this type of BMS features. We have proposed rigorous proof of convergence of the state energy parameters. And as I was saying initially, this is done for a fresh battery when it comes to have reused battery for example that we get out of the retired EVs. There are other edgy mechanism that come into place and those edgy mechanism interact each other in a non-linear and unpredictable fashion. So one thing that we would like to do here is to add the official degradation mechanism and use also machine learning to get inside about how to rationalize those edgy mechanism from data. And later on one important things to do is to test this algorithm on real data, field data to test the effect of effectiveness. And the other challenge that we're working on is the efficient deployment of this algorithm on large scale battery pack. So where we have hundreds of thousands of cells. So we thank you here for your patience. Yeah, thank you Simona. Maybe Simona will just take one question for the time consideration. You're describing this electrochemical base model instead of circular model. Then one question coming from audience is how much computing power is necessary to do so? I mean, how fast can it done? Can it be done, I would say in the realistic sense and the BMS on board in the electrical car? Right, right. So that's a very important question, right? That's the real question. Can we run in the real time? So there are more, one PDE, if you use a single particle model, you use the charge conservation on the electrode. In the solid phase, you have one partial differential equation, one per each electrode. And if you go ahead and discretize that, you might end up with maybe five ODE's for each electrode of maybe four. So in ECM, you have a two or three or differential equation. So you have a slightly higher number of ODE's. Yet what we have shown in the hardware in our lab is that you can really run the electrochemical models in real time. So that hasn't been done on the real vehicles and that's our next step. We really wanna prove that we can do it in real vehicle. But in our mind, that's our next step very next things to do. Now, that being said, we also need to acknowledge the fact that today we can rely on cloud computing. So we don't have to think about those algorithms to necessarily be run in real time on our ECU and the vehicle. They can be run through cloud computing we use with that platform. So probably that's the application, the very ultimate application that we are looking at. If we don't want to sacrifice the accuracy of these models. Yeah, thank you, Simona. I'll pass this back to Will. All right. Thank you E and thank you Simona for that very extensive talk. So now we have our second speaker for today at Professor Dirk Uwe Zauer from the University of Aachen. So let me briefly introduce Dirk Uwe. He is currently the chair of the electrochemical energy conversion and storage systems. And like Simona, he is a true engineer and electrochemical engineer and his work really reflects this intersection of science and engineering. I want to highlight that Dirk Uwe is also a beacon of inspiration internationally for academia, national lab and industry collaboration. He's a leader, for example, directing the Eulish Aachen Research Alliance and also in the leadership of the Helmholtz Institute Munster Network as well. And at least to us in the US, this is very inspiring to see how national labs and academia and industry are coming together. I know that Dirk Uwe also operates a very large experimental facility that welcomes industry users as well. In addition to his role in developing industry, academia and national lab collaboration, he's also a community leader. He is the editor of the Journal of Energy Storage that publishes many important articles in the field of electrochemical engineering. And then finally, I also want to highlight that Dirk Uwe is also a proponent and practitioner of technology transfer. I know that he is also has a number of entrepreneurial activities and also support his mentees on translating technology outside of the lab. So as you can see from my introduction, he is really influencing the field and the community in a multitude of ways academically but also industrially as well. So Dirk Uwe, we're very excited to hear your thoughts on how this is all coming together and sharing your learnings. Dirk Uwe, the floor is yours. Yeah, William, many, many thanks for the very kind introduction and it's really a great pleasure for me and a great honor to have the opportunity to share some thoughts with you. Not easy after this very broad and also deep dive of the Simona into all these things in battery diagnostics and modeling, but yeah, we will see that for sure. And I think this is also good news motivation but also problems and approaches are relatively similar around the world. Probably we're looking a little bit from a different perspective on this as William already has mentioned and yeah, I'm happy to share slides with me. I hope this works now. Yeah, so as I said, I'm coming from the Abit-e-Aachen University, this is a large engineering research university with about 47,000 students in bachelor and master courses and was very much dedicated also in research towards cooperation with the industry this was already mentioned. So we always try to find the balance between basic research but also transfer of results into applications. We'll shortly introduce what we are doing. So here we have the two chairs, Professor Ecklert Fiegemeier is with us since some years and we are looking into these fields of battery analytics and materials to modeling, lifetime prediction and then battery systems and vehicle integration including all these diagnostics and battery management systems and finally also the grid integration issues. So we have about 80 full-time employees in our group and the same number roughly in students doing the bachelor and master thesis or working as student assistants with us. Structures of German chairs are somewhat different maybe to those in the United States. So a little bit different sized in terms of relation for example among professors and students. But with a team we can cover a wide range of topics here in this field and let me say approximately 80% traditionally of all our projects are in cooperation with industry either via public funded projects. So we hardly get to the public funded projects without industry collaborations or by direct financing from the industry. So it was directed a transfer. So what we are just going for is to extend our activities especially into the field of these fields of aging, reliability and lifetime prediction. So we understand ourselves as an intermediates somehow between those who develop materials or production processes on the one hand and those who integrate batteries into the field. And this is why we're doing these things from deep understanding of the materials and we especially with our work we want to accelerate the time to market. So when there are new materials or new cell designs or new production processes, it is our task and this is what we really want to achieve to characterize these materials and cells in a short time so that we can give feedback and accelerate at the end of the day the overall development by fast, much faster feedback as we have it today let me say from pure benchmark tests. And this is why we combine electrochemical analysis down to the nanometer scale modeling along all the scales from the things we just have seen from Simona up to the applications in cars or in European electricity grids and with test benches. In our new center, CAR, Center for Aging, Reliability and Lifetime Prediction we bring these things together and we hope that we will rush in here in the end of May this year. Here we have about 5,000 square meters of the lab space and office space for about 150 researchers. And as I said, on the one hand, we really can look in deep into the materials also in situ in the battery cells, different size levels. And on the other hand, we also can put the full car into a temperature chamber. We have vibration shock facilities there, salty fog for environmental conditions. And this is very special. We have the third chair here by Professor D'Donca and he's doing the electronics. So we can combine also the look to the battery, to the chemistry and to all the electronic components, the battery management systems for sure from the hardware side, but also from the power electronics to see where we can find best overall system solutions, not just limited to certain things. Very important is here that we have a lot of test benches. We will operate you more than 4,000 test benches with currents starting from 20 amps to roughly 500 amps for different test benches for sure the majority is in the range of these 20 amps for the smaller cells. But we really want to go into statistical analysis and also, for example, if a production line is put into operation, we will be able to collect maybe every day 10 cells, have a semi-automatic analysis of these cells and put them for up to half a year into aging tests and this day by day. So to really monitor how the processes are improved or if some things are going into the wrong direction and to give this a faster feedback. And this is somehow how we represent this. So at the end of the day, we want to bring all the things, so we bring the things together to understand what happens if somebody accelerates a car. What does it mean for the current distribution within the cells and the battery pack? What does it mean to the distribution of currents in electrodes to concentration of ions and then down to what does it means to ion concentration in crystals and particles and so on. So to really understand what does it mean, especially for sure with regard to aging when somebody accelerates a car, but also the other way around. So if we have a new material, some colleagues come with a set of new materials that we want to try to characterize them in a way that we can tell them relatively early if the materials they are just developing will be a benefit for a certain application or if maybe the existing technologies already cover all what a new material can. So back and forth in this circle of innovation, this is our task here and this is all comes along for sure with lots of data collection, data management, analysis, machine learning and so on. And as already mentioned, we currently have four active spin-off companies who were founded by our former PhD students in the field of testing, modeling, diagnostics. So this company works for all the European-based car manufacturers in battery testing, for example, but also developed or supplying of models and diagnostic tools and consultancy. Ibu's plan is supporting public transport authorities or transport authorities in introducing electric buses. Safety is doing diagnostics, rapid diagnostics for battery management systems, but especially also for end-of-line testing in battery production. And the cure is looking into this from a data point of view. So they are collecting data and doing data analytics. So then for sure, this is a really cloud-based approach and they are collecting already today, more less than two years after they were founded from several 10,000 batteries, the data for many different applications, including automotive, but also ships or bikes or stationary storage systems. But before now, I would like to go into some details on modeling and control as well. Please allow me to pause for at least for a minute. We just are learning the hard way how small or otherwise seemingly big everyday problems are, but it's also time to reflect what and why we are doing what we do. We have a war in Europe and this is really something which is extremely difficult for all of us people are dying. Nevertheless, I think one of the big issues and Simona also raised this problem, the highly dependent on the fossil fuels from Russia, but also from other countries where we are not too happy to be dependent on is something we can defeat. So it was going for renewable energies for energy vendor and finally for sure energy storage, battery storage is a key technology to really introduce this. And yeah, this is main motivation for me since I'm now roughly 30 years working in this field of renewables and energy storage. So battery modeling and diagnostics, all the motivations and so on were shown greatly by Simona already. As we all know, batteries age also very differently. So this is why we need highly adaptable algorithms. If you look to the aging of cars as here for the leaf, we can see over time, there's a very wide spread in aging depending for sure on the operating conditions, but also as we all know, there are hardly two really identical battery cells coming from the production. And therefore, and this is the main part of our work, we would like to support the customers, the users to look into what they are really concerned of questions about range, safety, charging time cost, but also residual values. So most important thing probably in the future will be to be able to predict the remaining life of a battery at any point of time because if you want to sell a used car, the main value of this used car will be the battery and therefore this residual value is here of the major concern. And the battery, the challenges for sure also have been shown. The cells are different from the manufacturing process, sometimes very small differences in thickness of layers of internal pressures and so on already lead to significant changes. At least let me say in the last third of the lifetime of the batteries, this is what makes it so difficult that at the beginning of life, everything looks pretty similar, but not towards the end. Depth of discharge makes a difference, temperature, time, state of charge, the currents and also for sure, mechanical stress. And there's all results for sure in capacity fade and then power fade. And this is what the user really cares for. And questions are then coming from this. For example, how long will be the first life in a car? What is the residual value after the first life? How long could it work in a second life? For sure, depending on the operating conditions pretty much how reliable will the battery in the second life? So how often will we have to face some death of batteries or even maybe things like a fire and these things, lucky we are do not happen too often, but they are happening. And so we have to deal with this. And yeah, so questions also for sure, is there what are the applications for something like second life or is a second life at all something which is really interesting to discuss but maybe not part of today's discussion here. So when we look to the battery, we have to have the deep understanding and this for sure can be represented what is today called a digital twin of the battery. We all use these terms. So and this for sure is based on physics based modeling in different ways. We have machine learning algorithms for supervised learning and reinforced learning and parameterizing models or making also predictions. And so we need data, we need field data for all of this to verify our models but also to use them for sure as input into machine learning. And based on this, then we can do the diagnostics, the prognosis and also here the optimization of all this. Because at the end of the day for sure, we would like to recommend to those who build battery systems and those who operated what to do to have to achieve the longest lifetimes either by designing the battery packs and maybe even the battery cells in the most proper way but also how to adapt the operating strategies properly to get to the best what we can get from the battery. So the full cycle control here was the battery digital twin allows us in state monitoring, parameter identification, aging, estimation, capacity fade and power fade prediction. Also maximum power prediction is for sure an important thing, especially for example in hybrid electric vehicles where power is the most important thing and the optimization of operating strategies and as shown before all these things have to come together and this is what really is the great thing about battery research that is so much interdisciplinary and we need to really the specialists from all the different disciplines from mathematics, from mathematics, from mechanical engineering, from chemistry, from physics and also from electrical engineering here to really get an understanding of this very complex device called the battery. So this was already mentioned by Simona especially in the answer to the question at the end for sure the opportunities we have today significantly larger than maybe five, 10 years ago. On the one hand, we have significantly more powerful computing computational chips in the vehicles. So I started working on that asset batteries and if you look to the smart sensors on that asset batteries then we had to fight for every line of code for every variable because the capacity, the computing capacity for these very cheap sensors was so small, this has changed but in addition for sure we now have also the opportunity to transfer some of the things into the cloud and this is also part of our approach to take data from aging experiments which we do on test benches or from fields data and so we can then treat them here and we can learn things from these data and then we can exchange information back into the controller in the car and so this is how we can close this loop at the end of the day because for sure every vehicle then can send back their data their experience back into the cloud and is available then for all of the others. So I think this is most important for sure the cloud gives us more computing power but I think the highest value really comes from having such a huge data set of experience available from so many different applications and batteries so that for example somehow an error or an error problem develops further step by step by cell going to higher or lower voltages so the way how this happens could be compared with what I have in my car I'm just looking at with 100,000 maybe of other batteries where similar things may have occurred and I think this is what brings us really forward the biggest problem here to be honest is we hardly get access to the data from the car manufacturers at least the German or the European car manufacturers are not very cooperative in sharing their data from what they measure in cars until now but we are working on this. And then for sure we have the different ways for modeling as a basis for the prediction but also for the diagnostics and the car we have the physical based or physical chemical based models based on boiler followed Newman models we have these equivalent circuit models and we have newer networks at the end of the day they all have advantages and disadvantages for sure the physical interpretability is highest with the physical chemical models very important when we want to understand how to improve the materials or design of the cells the ability and self-learning ability for sure as highest with neural networks and implementation complexity and resources is lowest for the equivalent to circuit diagrams and the question for sure is how can we combine the best of all to bring things forward in the field and I would like to go into two examples here of our work and the one is physical chemical model based control for plating free fast charging so you will see several of the things Simona just has shown already for sure it starts to have a physical chemical battery model and I think in this audience I just can go through this without discussing the things really in deep we have to look into the structure for sure porosity and so on the mechanical data but then for sure into the diffusion processes lithium diffusion migration and the electrolyte but also for sure in the crystals to charge transfer reaction the diffusion in the particles themselves and finally we have two-dimensional model in this case here we're operating also three-dimensional models to understand for example how lithium is diffusing into a backside layer of an electrode or from the electrode overhang back and forth into the active material so all these things at the end of the day are responsible for improvements in capacity which we often see during rest periods in testing of batteries for example or even batteries just was laying around two weeks, four weeks or longer we suddenly have a significant different capacity and this is not due to aging but due to activation or loss of lithium into the overhangs or backside the coatings of electrodes so all this needs to be taken into account for the models which we're operating directly in diagnostics or in battery management systems they are for sure more reduced order models but important also as seen before is now to have the model on the one hand and an observer for sure to update the model and the parameters at any point of time that is clear a model without a parameter adaptation wouldn't help at all for what we are doing so the observation and updating is important and then for sure the thing what we can do with this is really adding a controller and in the example now of charging it is the relevant parameter for us from this model is for sure the potential of the anode of the negative electrode we all know that the potential for sure has to stay above zero volts and the potential for the reference to the lithium-dism ion reference electrode because otherwise we get this process of lithium plating so the position of metallic lithium which indeed results in very fast aging so and this is what we have done here we developed the model and we implemented it here on the battery management system we brought this into our test bench so the model is running on this battery management system platform delivers the value of the maximum current which can be applied to the battery without going into plating but for sure always to deliver or to take the maximum possible current to have the charging process as short as possible so this is connected then with the battery test bench and the battery test benches connected to the battery itself and we have for sure bi-directional communication necessarily because the data from the battery cell are going back into the battery management system here you can see such a charging process on the upper right hand figure the current which is flowing here so it's an approximately 8 amp hour battery cell and you can see at the beginning it can be charged at relatively high current but then towards full charging for sure was decreasing current this is the cell voltage this is what we can measure with the battery management system without any problem but which obviously does not hold the information we need because the information we need is this potential of the negative electrode and here you can see we have put the control value to 5 millivolts above zero and the controller keeps the voltage of the anode according to the model output for sure quite precisely at these 5 millivolts just at the beginning it's a little bit fluctuating here but this is what we have to measure and we have to measure it and take into account here as well so this is what we checked if this has a positive effect on the batteries or not looking to the batteries we check if the plating has happened by looking to the voltage during the following discharge charging for sure discharged and from looking to the voltage or to the deep analysis of the voltage curve we can see if there was lithium deposited metallic lithium and we can see this as an example here this is the standard curve here we applied and we increase the current rate by 10% or by 20% we can see that the curves are obviously changing and this is a clear indication that we have the battery for a longer period of time on the lithium potential and therefore this is an indication for having the metallic lithium here and the dotted line here is a constant current charging in addition to the fact that it has an impact on the battery and even though these are just 35 cycles we can see that if we go with a constant current charging so without control by our algorithms on the battery management system we have a capacity decrease following the dotted line here test has been done by zero degree C and if we control the voltage we can see that the voltage is just shown the aging is significantly lower because we can avoid the plating and not shown here still we were faster in recharging the battery with the controls in a controlled way and we saved a lot of time and this is for sure exactly what the common effect is now looking for so this is the biggest question we face at the moment is the use of the battery and the battery is also a big issue with the auto-motive and the emissions with the automotive industry because I want to have the cars charged with 250 kW and the question is how to do this without harming the battery dramatically and the second example the parameterization point of view, you know, that these models we just have seen also with Simona and say, I have just shown, they have a huge number of parameters. And so the first challenge is for sure to understand which of these parameters really have an impact. So where we have sensitivity on this or not. And here are several of these parameters listed here, which are necessary to describe a battery cell. And within the last 10 years, we have developed laboratory experiments to determine each of these parameters for any battery we get into the lab. So any type of commercial battery cell or also batteries coming from research projects, when they come in our lab, we dismantle them. And then we have measurement procedures to determine the parameters of these models, should it be porosity or diffusion or conductivities and whatever. And different procedures are used here. And at the end of the day, the question is, can we do this also in a different way to determine these parameters, at least if we have a battery cell operating. And this brings us to the question, to what extent we can also determine these parameters for machine learning or artificial intelligence. And for this, as also shown by Simona, it is necessary to look into the sensitivity analysis. So we apply, for example, via GP driving cycles to electric models. And from this, we generate battery currents, profiles in a huge number under many different conditions for different batteries, different chemistries and so on. To see at the end of the day, which of the parameters really highly sensitive, and which of them has a little sensitivity. Because obviously, when we want to have self learning models running in the car, we cannot learn 30 or 35 parameters just from measuring the current voltage and temperature, maybe impedance in addition. So we have to concentrate on those parameters which really have a high sensitivity towards the lifetime issues we're really looking for. And when after having done so, we can apply then machine learning algorithms. So we optimize or we determine, first of all, the most sensitive parameters here in the first optimization step or in the first iteration process step. And then in the second step, we go also for the minor relevant parameters as long as we are not in the car. So when we have sufficient computing time and lots of measured data available and so on in the car, then the step one is what we can apply but probably not step two anymore. So this allows us then based on so and the important thing is we're doing machine learning here, but we're applying it to physical based models. So the outcome from the machine learning are still values with physical meaning. So these are dimensions or diffusion constants or concentrations or whatever. And whereas if we just would do black box machine learning with neural networks, we don't have physical meaning. We maybe can describe the performance of the batteries very well, but we hardly can optimize materials or structures when not knowing what the physical meaning really is. And we were able to show that with this machine learning, we could determine the parameters even more precise than with our laboratory measurements. We are faster in calculations. We have lower arrows in capacity related to parameters and in generally. So we can see that there's really improvements and it is for sure significantly faster and cheaper. But and so here's just one more slide showing the training is based on the discharge processes constant current discharge processes that's based on pulse profiles and also on driving profiles. And then we can see when we test all this, we see that the the measure measured results can be represented by the simulation based on the model with the parameters determined with machine learning very, very well. The root mean square arrows here for different types of driving profiles are in the order of 10, 12 millivolts and on the cell level for sure. And we think this is pretty good for what we need. So here's just another comparison here between experimental parameterization and data driven parameterization. And we can see that the data driven parameterization is giving better results than the experimental parameterization. And this is also not unexpected at the end of the day. But it's really a very helpful tool to have this now available. So to sum up the challenges for battery management systems and future functionality will not end with what we have discussed today. We talk about over the air updates, not only of parameters, but also of models. We will see new sensors, pressure sensors, ultrasound, online impedance spectroscopy, for example, glass fibers doing measurements in the battery pack or maybe even the battery cells. We have data driven models, neural networks, and also more efficient data transmission and storage. Over the lifetime, the European Commission is really forcing the common effect also to have a full set of information available from the battery from the beginning of production to the end of life in recycling, for example. And it is part of our work also to check what are the real meaningful developments. So and finally, if we now compare what I have just discussed. So the physical chemical models was parametrized, was experiments, the strengths of this for sure. And this is important. So we are absolutely sure the machine learning will not replace the experimental determination of parameters. They are an additional tool. Because if we just have fresh materials, so materials which not have been built into batteries, then we cannot generate data which are needed for the machine learning. Then we really go for this with our experimental analysis of the parameter values on the materials. And then we can do virtual design and performance estimations are possible of battery cells. And also we can do extrapolation to even untested operating conditions. It's more difficult to rely on machine learning algorithms for the extrapolation. And on the other hand, the physical chemical models parametrized with machine learning. This requires that we have the cells which we can test on test benches under different conditions or in the field. And what the big strengths is as I try to explain this, we still have a clear physical meaning of this. And we have a more precise representation of cells after the measurements. So if we have a full battery cell with us, then the machine learning is a very good tool if we just talk about materials or half cells or something like this. And if you want to really design virtually new battery cells from materials from the database and also for different sizing of electrode sicknesses, porosities and so on, then it is helpful to have the physical measure or the experimental measurements. Yeah, but this I really would like to thank you very much for the attention. And once again, it was a great pleasure for me to share some thoughts with you on the battery modeling and diagnostics. And I'm happy that we may have some time for some discussions now. Thank you very much. Dear Gruver, thank you very much for that comprehensive talk. We're a little bit running late on schedule. So if I could suggest we ask Simona and Osta to come, we can now just have the discussion. How does that sound to the both of you? It's fine. Perfect. All right. So Dear Gruver, let me begin with a question which might be a bit provoking. So your institute houses a respectable amount of battery testers. And industry has one to two order magnitude more per site for their operations that 100,000 battery tester is not uncommon. But everyone, at least I speak to always says we need more battery testers because we're making billions of battery cells and the battery tester is a smaller fraction. So I would like to ask you, Dear Gruver, what has limited the effective use of the battery cyclers? Is there a way to make use of them more efficiently? And then secondly, what is the bottleneck to increasing the capacity of battery testing? So it is not bottlenecking the R&D process and the manufacturing process? Yeah, that's really most important point you make. So, and this is also one of our tasks, the question, how can we accelerate the testing on test benches? And when we look to how battery cells are tested still mainly today, you apply to one cell a certain current, so typically one constant current at a certain temperature and you go for a certain depth of discharge. And then you may take two or three other cells doing the same to have minimum statistics of three cells, which is from a statistical point of view, not enough at all, but nothing more typically happens. And then you apply with higher current to the next cell and so on. So the first question we raise and what we are working on is how can we really go for more complex testing profiles? How can we combine in one test profile different current rates, different temperatures, taking differences? And here we work also with our colleagues from mathematics to get the info, still the good information, because we want to know what a certain current rate does to the battery. We want to know what a certain temperature does, because otherwise we cannot parametrize our models properly, but we are sure the future cannot be to stay with applying just one set of applications to one cell. We have to bring various multi-stress profiles for the batteries. This is one side, but then it doesn't end. For example, until now I would say the number of tests which are done, for example, where we test electrical profiles, temperature profiles, and mechanical stress profiles at a certain point of time. And this will become much, much more relevant when we talk maybe about solid-state batteries, will be superior lithium electrodes. I fear that they will be extremely dependent on the pressure we apply to the battery cells and so on. This needs to be combined on very expensive test benches. We have to understand what a certain stress does to the battery to separate these stress or the effects from the stress later on. Otherwise we will die in battery testing. Dr. Grover, I think you probably cannot answer it, but just to inspire the researchers in the field, how much improvement do you think there is to be made in terms of the efficiency of battery testing today? How much headroom do we have to speed it up? So our vision, our vision and mission for what we are going for is to be able to give feedback here to those who produce materials or battery cells for sure with a different level of accuracy after four weeks, after eight weeks, after 12 weeks. So within 12 weeks we must be able to give a clear statement if a certain battery will have the ability to fulfill the requirement of lifetime in a certain application. Otherwise we can't speed up enough and you know at least in Europe we are well behind the Asian battery manufacturers so our industry also have to close a gap and if we can't help them from a scientific point of view in accelerating these development processes we won't be able to catch up. But I'm pretty optimistic that within 12 weeks by combining very specific tests on test benches which go for example, the different tests have to concentrate on different aging processes like plating, like current distribution in the battery cells, maybe corrosion or whatever. Then the post mortem analysis was really in the in-depth analysis of all these processes to see already at the beginning even though it does not have an effect on performance yet if we have ruptures of crystals for example or what kind of solid electrode interface is built and then the modeling to extrapolate this all to the real world operating conditions of 10, 15 years in the field. And this is I think what the research community has to bring together and with the formal point of view the holy grail of battery, lifetime and reliability prediction and the testing is part of this but it must be much different from what we are doing today. Degrove, I completely agree and also resonate with what you and Somona emphasize today is what we need is more actionable insights coming from the testing just certainly knowing whether it's going to pass or fail a certain criterion it's utterly insufficient because it doesn't tell us as battery engineers and scientists how to improve the materials how to improve the cell so I completely resonate with that. Maybe also let me take some of the points you and Somona make and expand on a bit. So both of you talk about battery management online estimation it is already a hard problem today given the complexity of the battery pack the number of individual cells involved but we're also seeing increasingly a trend in order to increase the flexibility of design that multiple chemistry two chemistries are now being introduced within the single battery pack and of course this increases the complexity of the power electronics and increases the need for even a more powerful BMS system and the coupling of the two as well so I was wondering if the both of you can also comment what are the opportunities and also challenges when we start going to dual chemistry battery pack and what is the trade off here between the flexibility of design versus the complexity of implementation for such a system maybe I can ask Somona to weigh in on this first. Sure um no no that's uh that's a very um important question Will I wanted to add something to the previous question you asked just about testing I wanted to say that when it comes also to test reused batteries for second life application that increase the testing capability needs because we know less and less we know even less about those batteries than we know about fresh batteries so I want to say that test is going to be a huge deal especially for reusability and uh re-purposing of those batteries when it comes to integrate different chemistry in a battery pack I would say from a BMS standpoint the um integration is is uh is the big challenge there integrating uh the electrical behavior uh in a way that you you get the performance that you want to from there from the pack and uh as a strategy for uh design and BMS in that case you will just want to start with very simple electrochemical and circuit models uh you don't want to start with electrochemical things especially I don't know there are combinations of lithium ion batteries super capacity or lithium ion battery and the lead acid batteries or differently to lithium ion battery chemistry so LFP combined with NMC so um the challenge is really to understand whether or not the integrated system will last and perform as desired over a long period of time and Tersi really should go there to uh to to understand what the um you know things needs to be tweaked and perform you know improved and changed um and the other thing is also the all the control strategy underneath the modeling that needs to be revised uh you know there are different time counts and that those energy storage might have and so how to account them and how to integrate them properly uh aging uh will have different trajectories so if you combine for example LFP and NMC we know the LFP can withstand very high C-rate of operation as opposed to NCA or NMC so uh how do we you know split um that load in an optimal way so there is a an energy management problem that you can establish over there that you can do it through some active uh control so power electronics that helps you to um you know split that power that load in a in an optimal way and in order for you to do that you need some uh you know good modeling to start with especially when you come you know aging modeling electrical model and the thermal also uh model response. Thank you Simone. Sir Gruber your thoughts on the dual chemistry? Yeah maybe maybe I can add from it was an example from stationary batteries for sure we have looked into this also for mobile applications but we operate since roughly six years a five megabit battery and which is operating in the German grid and grid control so we're taking part in the auctions every four hours to deliver grid services to keep the frequency at 50 hertz and in this system we have five different battery technologies so we have the two lead asset batteries and we have LTO, LFP and NMC technology and here we we're doing exactly this we're deciding at every second how to distribute the energy to the different battery technology systems in most most times in frequency control you're having five or ten percent of peak power and so if you need one percent power for sure you have to take all the batteries at the time but this happens rarely almost never and so what is used is well typically you have a relatively small thing and this is here we really have we have a digital twin for all the batteries representing on the one hand their efficiency including what does it mean for the cooling of the system and into what is the efficiency of the inverters which are connected into transformers and on the other hand what are the aging what is the aging so in every second we calculate the marginal cost for the operation of a certain battery and then it is something like an internal auction and we say okay now this battery will will be on online and the others go down or step down or we bring it to two three strings or whatever we have 10 battery strings which we can control individually here and this is I think this is a key you need really understanding of the of the aging and of also of the other things and maybe in the pure battery electric vehicle I'm not really sure that we will see too many different technologies with large batteries if you go to 70 or 100 kilowatt hours these batteries have sufficient power anyway so the combination of different chemistries is more interesting in small batteries where power and energy ratings are of interest but I'm absolutely sure that it will become relevant also for these type of applications because probably we will maybe change over lifetime individual modules and then maybe even the chemistry remains the same we will have in one battery pack old modules and we have new modules and the new modules maybe even have different chemistry because they are not produced in the same way anymore so I think the question you raise is most most relevant but in this point I'm quite really optimistic that we can handle this because this has been shown and can be done but for sure every more understanding of the aging and this is again something where cloud services will help us to see what has what what has happened in other systems and take these informations into my specific system a digrover is impressive that you have already been practicing this five chemistry system in a life setting this is this is very very exciting we have a couple minutes left and I thought maybe I would end with yet another provocative question and and this one is a bit close to home I would say and I will begin since I'm a professor I shall profess and rent a little bit you know one of the great things in the couple of years leading to today with battery informatics is that more and more battery data being shared the groover Samora and myself we have been sharing data which is great but I'm also seeing a concerning trend in which folks are beginning to just throw you know every possible machine learning methods to interpret the data and of course not surprisingly everybody gets great results so I think there is now this movement in the field to really exploit the most out of these public data set but of course these data sets are very small fraction what's available out there and this can often lead to over interpretation over fitting and under delivering the models so I'm a bit curious on what your advice to the community which are just now getting into machine learning and battery informatics what are some of the best practices to avoid this pitfall of just discovering unimportant and unphysical trends which are not really translatable to the real problems that's been deployed in industry maybe Degro I can ask you to comment on this I know you probably have some strong strong thoughts on this as well yeah I just want to try to ask if you asked me as an editor of a journal getting an numerous number of papers on similar things I don't have I don't have the answer on this I have to say I sure have many thoughts about this maybe machine learning is an example but even going one step before yeah let me say the traditional diagnostic algorithms what I see are a huge number of papers all dealing with some equivalent circuit diagrams and some observers to go for this and from my point of view in the community we have to be much more critical about looking into these things if I see an observer or a diagnostic system which is then verified at 20 degrees C for a new battery and for a well-known system you can say nice but this is pure academic yeah so if the algorithm is not working on the temperature range of minus 30 to plus 50 degrees C and if it's not working throughout the lifetime from a fresh battery to an age battery we should not really accept this as as a new finding yeah we have to define quality controls to really say now it is approved that what you're doing is is really good and also in machine learning I often see papers where it is said that I now have precisely determined that this type of algorithm is the best one and the other one not perfect and my first question is always do you really have optimized all the internal parameters of this machine learning algorithm or optimization algorithm can you really be sure that you have got the best from this if you make such a statement which of these things is the best and the answer in most of the cases is probably no yeah and therefore I think we have to be more critical about what we're really accepting also in the science community and we have to draw exactly the questions you just raised to do these things to find out what the real big things and not just publish things to publish another paper yeah I think Doug Groover I think that the challenge I personally see here too is it's very easy to do these analysis and just download the data set and run the analysis so the barrier is getting lower and I completely agree with you that more intense innovation is needed here and not just trying the same methods over many minutes or sorry to to profess a bit here but I'm glad Doug Groover you share the same concern as a no head to chair as well absolutely so Simana what are your thoughts on this absolutely I'm completely aligned with both of you and you know in fact in my lab I wanted to start saying that there is a positive side of this there is more people interested in battery people who are from computer science really are skill in machine learning and they they they approach this field you know through their knowledge but they don't know much about battery so it's it's good in one way and we'll see this every day we'll write we have so many people students wanting to start doing machine learning for batteries but they don't care about your physics so the electrochemistry understanding the underlying governing principle of these devices which is a no a no no for us right so you really need to be wanting to know how the system works because otherwise you're not going to be able to really provide any novel inside or novel discovery or anything like that so for us for me my group it's it's mandatory to have an interest and a passion for how this device works in all its form single cell manufacturing a pack or or whatever setting you want to to look at and agree with you there is so much shallowness around sometimes in in a a plan a machine learning algorithm and claiming that it works you know as we said also from professor sure you know you cannot extrapolate the behavior the performance of those algorithms across other chemistry or you know different data set and so forth so those algorithms taken by themselves can be promising but they cannot be seen as the solution to bms problems for example but the combination of the data driven or data analytics based method with electrochemistry with control theory with optimization that is where we should be really working on so the combination of different field is what really will lead to a big advancement improving in the field and do you have a paper where you lay down different types of structure right or how we can combine those in different domains and that's what we really should be working on with all our students and the whole team industry. Simona and DeGrover thank you both very much for this brief but insightful discussion and for taking the time early morning for Simona and late afternoon Friday for DeGrover to share your thoughts with us I and Ian I truly appreciate it so if I can have the closing slides please Evan thank you so this concludes the winter quarter of the storage ex symposiums so we will return in April with more exciting speakers so please stay tuned and with that I'd like to thank everybody for listening in today and look forward to seeing you next time thank you