 Well, thank you. Thank you everybody for joining in for this talk. So I'm going to be speaking about some work which I've been doing with my collaborator, Owen Maiden, who is another open force field researcher, and we've been really trying to do a data-driven study to try and elucidate which types of physical properties should we be optimizing the non-bonded van der Waals interaction parameters of our force fields against to yield the largest improvements in terms of force field performance. So to begin, I just wanted to give an update on the infrastructure, which is not only going to allow us to do this study, but to do many studies like it, because we're a really quite exciting point where we can now almost routinely refit the non-bonded interaction parameters of our force fields against even larger data sets of physical property data. And this is really made possible by this seamless integration between force balance, which is our fantastic optimization engine created by leaping Wang and his group at UC Davis, and the open-fifth evaluator framework, which I've built as an automatic scalable framework for estimating data sets of physical properties from molecular simulation, as well as their gradients, which force balance can then consume to take informed optimization steps, leading to, with almost no human intervention, these nice refit force fields. Additionally, in addition to just having this refitting pipeline, we can now routinely take our refit force fields and then use the evaluator framework to benchmark them against to gain large test sets of physical property data, gain insights into where the limitations in our force fields may be, and then take those insights and learn and make data-driven decisions based off them to decide what changes might be needed to make to our force field functional model. Maybe we don't have enough chemical environment types in the right area. Maybe our training data set is deficient in certain areas, but we can now do these kind of cyclical self-consistent optimization epochs with our almost no human intervention, with the only human intervention really being required to use the training and the test sets, but even that we're working to automate. And so the study that I'm going to talk about now is really the first major study that we've done using this automated fitting infrastructure. And so we've been trying to determine which types of physical properties are most informative to train the van der Waals non-bonded parameters of our force fields against. So historically, most force fields have taken the approach to train their non-bonded van der Waals parameters against only pure properties. So pure liquid densities to try and get information about the relative sizes of the molecules, pure entities of vaporization to try and get some information about the cohesiveness between energies. But while pure densities are kind of routines trained against, and until these vaporization are challenging for two really main reasons. For one, it is really, really difficult to find diverse and especially accurate data sets of entropy of vaporization measurements, especially data sets which are permissively licensed and open for us to use. So there's not much data to train against. And actually in terms of calculating these properties, we know that our fixed for our fixed charge force fields can't actually or can't at all capture the change in polarization when a molecule goals goes from the liquid to the gas phase. So if we're training against this, we have to either hope that this this contribution from the change in polarization is small. But if it's not worth essentially building electrostatic contributions into our van der Waals parameters, which we don't want, all we have to try and correct for this. But this is just another potential source of error into our calculations. So we as a consortium have been really interested in maybe moving away from pure properties, or maybe at least finding properties which are complemented to them, especially a mixture of properties, properties of systems of mixtures, especially enthalpies of mixing, binary enthalpies of mixing as a potentially better replacement for entropy of vaporization as it doesn't have these limitations, and also binary mass densities of liquids, either to complement or possibly replace these pure pure liquid densities. And we're excited about these properties for a couple of a couple of key reasons. For one, there is a significant amount of mixture data to available to us. There's hundreds and thousands of these mixed properties available in the NIST thermal archive of of thermal physical properties, which we can automatically pull down and pure rate in machine readable formats. We think machine mixture properties with mixed properties, it should also be easier to build training sets which are targeted towards reproducing well specific interactions in an almost modular way. Consider if you want to capture, say, alcohol ketone interactions with mixed properties, all you'd have to do is find a mixture where one component is an alcohol, the other component is a ketone, and on mixing use and capture this interaction. Whereas with pure properties, you're going to have to find these weird molecules which can take more and more different functionalities. So it's a lot more difficult there. And I think one of the big potential benefits, and this is mainly just a hypothesis at the moment, but we think possibly improving on these mixture properties, especially entropy of mixing, may correlate with also improving our performance at predicting things like binding for energies, salvation for energies, partition coefficients, things which rely on a lot of different interactions between components which which entropy of mixing will inform us about. So to kind of investigate this problem, the approach we took is to take the open force field 1.0.0, essentially the first parsley force field, and then refit a select number of its van der Waals parameters against different, against essentially four different training sets, where each training set contains a different set of combinations of the properties that I just mentioned. So one containing only those pure properties is kind of a historical baseline. One containing only those mixed properties, which I spoke about of being of interest. But also then combinations of these pure and mixed properties, with the idea that we're not quite sure if mixed properties are enough to constrain things sufficiently. i.e maybe we train on mixed properties and improve mixed properties, but at the expense of making pure properties worse. So we want to also see, you know, do we need to include pure properties to make sure things are kind of consistently constrained. For each of these training sets, we just focused on a limited set of functionalities for which we had the most diverse data available to us. So sets of alcohols, acid, zesthes, ethers, ketones and alkanes. But with the hope that any results gained from this study on this functionality set should continue to generalize as we move to broader, more diverse training and test sets, like what we required for when we try and refit for sage later in the year. So in order to do essentially these four unique optimizations, we need to build both a pure training set and a mixture training set. And then for those training sets, which contain a combination of both types of properties, we could just combine these two distinct sets in different ways. Essentially the pure training set that we constructed had a total of about 56 pure data points, and even a mixture of pure liquid densities and pure entities of vaporization, all the kind of ambient conditions, but for a diverse range of different components exhibiting the functionalities of interest. For the mixture sets, this contained about 200 data points total, so about an even split of entropy of mixing data points and binary mass density measurements, where each of these components was the mixtures in the set. And for each of these mixtures, we tried to choose data points for three different compositions, three different more fractions for each mixture at 25%, 50%, and 75% composition, again at ambient conditions. I think one of the main goals in trying to build this mixture training set is we're not only wanting to get a diverse set of molecules, but we really wanted to select mixtures which exhibited behavior all the way down from almost no deviation from ideality, i.e. where these mixing properties are mainly determined by the properties of the pure components themselves, all the way up to large deviations from ideality where in the mixture there's new interactions which weren't necessarily present in the pure components, such would be the case in say alcohol and esters, where esters themselves can't form hydrogen bonds, but they can with alcohols. And it's this region of large deviations from ideality that we think training against these mixed properties are potentially going to be the most informative. So then for each of these four training sets that we built containing these four different combinations of properties, we did our optimizations using our automated fitting infrastructure, and here I'm showing the objective function of the fit as a function of the optimization iteration, and one can see that all four of these, the objective function decreased almost immediately by a function at a factor of about two to three-fold. So it does definitely seem like our optimization is pushing things in the right direction. One can also look at the RMSEs of the different properties at the beginning in the end of the optimization and almost universally across the board, things improved, with the exception of maybe one outlier which I don't necessarily want to go into in this talk. So the question then is how well do these results generalize to broader test sets of molecules? So to kind of test the generalization, we built a significantly larger benchmarking set than the training set. It contained about a thousand total data points. It contained again pure liquid densities, and the piece of vaporization for all of those different functionalities that we were training against contained a significant amount of mixed data points, so entities of mixing, binary mass densities again, but also we included binary access and all the volumes just to make sure we're getting the difference between the pure and binary densities, hopefully correct. And again for each mixture included, we looked at three different compositions and kind of ambient conditions. I think the main deviation here between the test and the training set for the mixed properties, as well as for the mixture training set, we only included say alcohol, alkane, alcohol ester, ketone ether, alkane, ether and alcohol acid mixtures, but in the test set we tried to contain all different permutation pairs of these different functionalities, including things like alcohol, alcohol, ether, ester, even though we didn't necessarily train against them just to ensure things were generalizing. So we did the benchmarking, we spun up our automated benchmarking infrastructure. Here is the results of that. So here I'm showing the root mean squared error of each of the different properties that we were benchmarking against, where each of these colored bars represents the results of benchmarking this property against this force field, which was optimized against this kind of unique combination of properties, where this red bar was a benchmark done upon the initial open FF 1.0.0 force field as kind of a benchmark baseline. So any improvements relative to this red bar means things got better essentially. So what we can see from this is that almost universally across the board, things did improve regardless of the types of properties that we were optimizing against, but looking at this a bit more particular. So this far right bar, which I believe is light blue, was a benchmark done against the force field optimized against the only WRO properties. So one did see some improvement in the pure properties. These are 95% confidence intervals, but so we can't really say improve too much because there's no statistical significant improvement here when training against only WRO properties. We did see improvement in the binary mass density, but we saw almost no significant improvement in the entries of mixing and actually a significant degradation in the excess molar volumes. And again, you know, going back to what I was saying at the start, we were hoping and we kind of wanted to see improvements in entropy of mixing because we think this is going to be correlated with those properties more of interest. So compare that to these first three bars, which were benchmarks done against force fields optimized against mixed properties. We do see significant improvements to the mixed properties, even though we're benchmarking on different type of interactions that we didn't necessarily train against, we do see significant improvements to the binary mass densities and we don't necessarily see at least improvements, but we don't see degradation at least of the binary excess molar volumes. But I think one of the key things that we do see is actually if we look at their performance of predicting the pure properties, at least, again, not too much significance here probably due to the small size of the pure property benchmark set, but we don't degrade by only training on mixed properties. We don't degrade their ability to predict the pure properties. So it does seem like we can get away with only optimizing against mixed properties without necessarily having to worry too much about things being being under constrained. And again, just highlight the reason why we think we should be quite excited about these improvements in entropy of mixing is because and we're testing this hypothesis now, but we hope improvements here should correlate well onto some of those other mixed alike properties which are the most of interest to us. So one can also look at the r squared, so here exactly the same kind of plot, but now on the y axis we've got the r squared and not too much extra information here. And theory main takeaway is actually when we benchmark the entropy of mixing against a force field optimized against the mixed properties, we did see significant improvement, a significant improvement in the correlation of these properties. And we can dig into that just a little bit deeper. So here on the x axis, I'm showing the estimated values of the entropy at the binary entropy of mixing for each of the different force fields that we're benchmarked against on the y axis, the experimental value for the binary entropy of mixing. And if we look at this left most plot, which was the benchmark done upon the open ff 1.0.0 force field, we see these large clusters where there's a significant systematic error in certain mixtures of functional groups, where each of these colored groups, colored dots, represents a different mixture of different functional groups. So especially alcohol and esters, alcohol and ketones, and possibly even alcohol and alcohols were poorly represented in the initial starting point. And if we look at this end plot, which was the benchmark done when we'd only retrained on pure properties, and we really don't see any help in improving this systematic error, or at least not much. But compared against when we optimized against the mixed properties, perhaps not too surprisingly, the systematic error almost completely went away. And the reason why I'm highlighting these orange green and red dots in particular is because these are mixtures which show significant deviation from ideality, i.e. there is interactions present in the mixture, not there in the pure. So including these mixed properties optimizing against these mixed properties do seem really important for capturing those deviations from ideality, and we do think this should help improve our force fields significantly. So just to wrap up quickly then, I really want to emphasize again, we've got down with this fantastic automated britting infrastructure, both on the QM side, which I've shown here, but also on the non-bonded side. These kind of studies would have been so laborious and human intensive previously, but now we can do them almost effortlessly, when most of the human time just goes into building the training and the testing sets. So this infrastructure is allowing us to do data-driven studies into force field science. We're now seeing from these studies that actually mixed properties do seem to be a good target to head towards for future force field optimizations, especially it seems like we can replace this problematic enthalpy of vaporization with our enthalpy of mixing measurements, and not degrade, but actually hopefully gain performance. And the last thing, and the kind of missing piece of the puzzle, which unfortunately the calculations haven't finished in time, but we're now doing these benchmarks against expanding the benchmarks to include a host guest binding affinity to really try and test this hypothesis that improving these mixed properties in particular enthalpy of mixing do lead to improvements in these host guest binding affinity and other mixed alike properties of interest. But we should be sharing those results in the next couple of weeks. So as everything that we do in the consortium, all of the input analysis scripts for the study is on GitHub, the frameworks that were used are available on GitHub. And just to finish off, I'd again like to thank my collaborator Owen Maiden for his his help on this project, Philippine Wang for his substantial efforts in helping to create integrated evaluator into force balance, and also Christopher Bailey from OpenEye who has given such a significant insight into helping to design design and analyze this study. And then of course everyone for their attention. So thank you so much and if there's any questions, I'd be happy to kind of answer those now. Hi Simon, this is Jamshad. Hi Jamshad. Really good stuff. Nice to see you again. I was just thinking out of this, there's a really good scientific question that could come out nosefully easily. And that is that, you know, the potential functions particularly with van der Waals parameters, we did the parameters, I guess originally were optimized on pure systems. And then we have these built-in Lorentz pathological mixing rules. Now, you know, I'm not quite certain what the history is to the mixing rules. I've intended to look into them at some stage, but I've never really done that. But there is a good rationale for them or whether they are empirical. Now, what this could in fact is to actually determine what the limits of those mixing rules are. So if you did actually focus on the pure properties and then see how well we are able to reproduce the mixing data. So my point is that one, there's two roots here. One is to, of course, to optimize the van der Waals parameters themselves. And the other thing is, or maybe one could do both. And the other thing is to actually focus on the mixing rules too, that one could perhaps either derive a better mixing rule. Because at the moment, I think those two are also parameters. Yeah, so I think they're absolutely spot on with this. And this is an absolutely fundamental issue that we want to look into the future. Like you say, we know these mixing rules have been chosen historically. I think the mixing rule for the sigma parameter was chosen based on hard spheres. And I think the epsilon mixing rule parameter was somewhat chosen to reproduce effects of electronegativity. But as a consortium, we definitely want to look into this and explore what mixing rules we should be using and what impact they have. So like you say, we could potentially do a study such as this to explore the question. And that is definitely an avenue. But we're also looking at more quantitatively rigorous approaches to also study this. So on Friday, my colleague Owen Maiden is going to be speaking about Bayesian methods to try and quantify, to kind of put a number on, you know, what's the performance of these of these different mixing rules and other various aspects of the force field for a given data set. And so that's also a potential study and route that we're looking into, although that one's a bit further off. Thank you. So have you, do you have thoughts yet on places where it looks like you especially will need to couple charge fitting into this? I guess as BCC fitting comes online or is that a little too far ahead to see yet? Yeah, I mentioned when we did the optimization and actually we saw this in the benchmarking study, when we kind of split up the performance into, you know, what kind of environments were there, almost universally across the board, we saw that the performance of predicting ketones density, either binary density where ketones were in excess of pure density where ketones were just on themselves, almost universally degraded. And it was just ketones. So we think this may be linked to maybe the van der Waals having to accommodate and kind of correct for deficiencies in the charge schemes there. So I think this is one of the avenues where we think this might be might be important. But we're still in the early stages of looking into that. Great, thanks.