 So invited speakers generally told us we were good organizers up to this moment in which we prepared the panel session. I mean I'm a bit fond of this format in which I hope you will have all the opportunities to exchange questions mostly related about where the field is moving. And yet it's still an occasion to debate and discuss publicly about ideas and thoughts you might have had during these three days. I'll be the moderator and you will find Alessandro Lio, Christof Schutt, Francesca Grisoni, Sasso Gerovsky, Andrea Nellie and Julia Westermeyer in the panel. And they've been gathered because of the inexpertise which is mostly related not only of course but mostly related to working on the design of molecular systems. So I'll break the ice with one question and then I hope that the stage and Zoom will contribute with even further questions. And my question is if you had to identify one problem that needs to be solved to push our community further and be even more helpful in defining better drugs or better molecular systems with greater functionality, what is this problem and how do you think this might be solved in a few years? As I mentioned that's the moment in which go ahead, the bravest first. Okay, since I am the least material scientist here, to me it appears that quite an important problem is to distinguish clearly between molecules and materials. Because my intuition is that there is more to materials than just the molecules. There is also the structure, the crystal structure, but I'm a computer scientist, so I don't understand fully the complexities of this. I just know that there is this difference and I think this difference should receive more attention and systematic ways to describe the structure or whatever other aspects there are to materials in addition to the molecules. So this for me is quite a priority and on the computer science side I don't think it will be that much of a problem to represent it, but it's very important to understand it and I have the feeling that this understanding is not really widespread. Well, so I don't know if I am able to identify a real big open problem which has not yet received a lot of attention. So there are many, many important things that have to be done in order to really achieve, let's say, a fully reliable tool like, for example, treating electrostatics or like dealing with polyatomic systems with very, very heterogeneous features. Well, all of these directions, I think there are ongoing advanced attempts, so I wouldn't say that any of these problems is in any manner literally speaking unsolved. So there are steps forward that have already been done, other steps forward are for sure necessary. So if I can make a generic comment on how the community is exploiting machine learning and artificial intelligence tool to solve like, yeah, for example, design molecular design problems, my personal impression is that the level of complexity of the machine learning tools which are used, they say in our community still lags behind what is out on the market now in other fields, like, for example, image processing or natural language processing. So the thing is that in these other fields, well, the idea of using a two layer perceptron fully connected is something that, yeah, people would look at you like if you were crazy somehow because now there are much more advanced tools. Like, for example, it is becoming clear how the role of depth of deep neural network for learning complicated features is absolutely essential. It is also becoming clear why width of the representation is important. So all of these things on one side, it is becoming theoretically clear why you should do it. And on the other side, people are using these things more and more successfully. So here I actually see a lot of for improvement by, let's say, hiring from other communities the tools that are continuously coming out and that are in fact more and more advanced and basically to translate this tool for applications in, for example, in material design or molecular design. So I think what we've seen at this workshop is that a lot of the previous research, so in what I'm doing most of the time, which is neural network potentials, was dealing with how to really describe local environments. And I consider this problem, yeah, not perhaps not completely solved, but really almost solved with what we've seen here. I mean, all the solutions start to converge as we've seen with these generalized ACE techniques, for example, because and if you look at at the equivalent neural networks like MacWhip or Paine, you see some advancements in the technical side with Aldeco where it's about scalability for large systems and so on. But basically, all the solutions converge in this to, yeah, to tensor products in the end. And I think that this is now going towards basically like you would say in an image context where you say I'm using a convolutional network. Now in an atomistic context, you would say I'm using an equivariate network and then you use these layers that have some details that make some difference, but it's nothing completely crazy. So what I think is the next logical step is to look at more of the semi local and the long range part and also something that only few research publications have dealt with previously in this field is different charge systems, spin states and excited states. So I think this is really an important topic in the for the next years. Okay, I can give a perspective more from the try discovery domain. And I agree with some of the things that were said, like for sure we need better representations for certain types of problems like what happens when you try and improve formulations. It's really difficult to represent formulations in such a way that we could play around with it and learn in an end to end fashion from those. But I think that what we really struggle with here is mostly to innovate. Like whenever you want to for example in the design domain, whenever you want to produce something new, something slightly or quite different from your training sets, you struggle. So I think active learning is really promising to help us generate our own data. And I dream of like fully integrated platforms where you can quickly iterate over your data, you you launch your calculations there, the molecules get synthesized and tested, then you get it back so that we can generate the data we need and slowly moving out from our training sets molecules. But also I think along these notes, we could also try and improve the out of domain generalizability of our models, because otherwise we are stuck with what we have and with the diversity we have in our training set. So this is where I think we should try and move. And definitely of course it's always important to connect with the domain experts. So I think explainable AI will play a big role for us to to get accepted by the experimentalists and to connect and learn from them and have them learn from our our models. Yeah, I think that's it. So yeah, I also want to add here. I actually agree with all of these points because they are very good. For instance, we need to distinguish molecules and materials and then we need to find ways to describe both of them. Like how can we enable heterogeneous catalysis? And as Christoph mentioned, it's also extremely important to somehow investigate different states, different spin states. How can we make use of light? Because there are a lot of materials who can harness light. We could use this energy to tackle most of the most pressing issues in our world currently. So I think this is one of the most important topics to tackle. Well, last point, I'm going to follow up a little bit on what Sassan Alisander just mentioned. But all all their points I think are perfect starting point for the discussion. But my I resonated particularly with the two aspects that were brought up. Because from Alisander's perspective, we're a bit lagging behind the development that has happened in other domains of field. And and we can catch up. That's for sure something. But what what I also observed from my point of view was the fact that we could not leverage the same synergies that the other fields could. They had a shared starting point, perhaps images or representation that they were no doubts worked for their exercises. For us, instead, I felt and this is something I think that expands a bit on what Sassan said before, that the it was not an obvious solution or an answer. What kind of representation? What do we even mean by material or molecules? And that's something that we're slowly getting the words to by fragmenting it, right? We've done a divide and conquer approach to how we're going to tackle the issue. And we started deciding that there's going to be a local information that's going to be a long range information. And we start from the local things. And it looks like we're doing well, but we're progressing at a slower pace also because we're still following this difficulty in distinguishing what is what defined the solid and what is what defines the molecule. So for sure, by continuing this trend, it will make sense to go long range to complete this ladder to complexity. But I feel the answer could also lie simply rethinking deeply on a more generalized way to look at something that is made out of atoms in a hierarchical fashion, which is something the community with more biological systems is used to work with, because in the end, you have structures and substructures. And from a materials and molecular perspective, we're still fragmented the data on the condensed matter. You look at local densities and then a carpet that contains everything else. And from a molecular perspective, there are times to do hierarchical stuff. But I now speak out of my frustration. I sit in a condensed molecular community, and I don't benefit from either because we are long range, we are a super molecular for some aspects, but we are also short range. And there is nothing less tailored for this. And I am not sure whether long range will bring us there in the shorter term. It feels like we need to have in the mindset a hierarchical fashion from the very beginning and reconsider the whole way in which we define anatomic or a molecular entity. So that's my little point. Thank you very much for all of this reflection. I mean, I totally agree if I can say a bit mine, but I'm the moderator, so I can. Yeah, integrating the way in which the final molecular material is done is going to be key. And I think it touches all the points you made about the complexity, formulation, also the question by Nui about the way in which the molecular crystals have been made. So we're going to predict not just the property, but the process to go to the properties. And in this regard, tomorrow panel session will be also related to energy materials. And we will see also their opinion on that. But at this point, I would open the floor for any questions either targeted or to the broad audience and whoever wants answers. So please feel free to break the ice also from the audience. Kevin, may I have a first comment? So I was very glad that this issue was brought up about the latest machine learning and other advances not easily making it into uses into material science because it is very, very strongly related to the point I brought up of having a clear understanding of what materials are and how to describe them. Once this point is sorted out, you let people from machine learning in the other areas stream in because they will understand computationally what the problem is. This is what they are brought up to understand and then they can attack it. If they don't understand that, you are limited to the kind of materials people that have learned a bit of machine learning. And this I think is a point to understand and is crucial for the development of the community. Yeah, I think, yeah, I fully agree that the problem is also formulating what we really want. And in particular now, after having thought about it a little bit, after the things that you said, I think that a challenge is a fully meaningful semantic description of a really complex organic molecule. So because if you think about it, if you have a complex organic molecule, imagine something with many aromatic rings, many links, many everything. So you think of course now we have features that are able to describe the local environment. But so the global connectivity, I know that there are tools which allow to encode, like do positional encoding, which is also aware of betapology. But so it's simply complicated because you can have branches, you can have loops, you can have everything and all of this must be meaningful, must be encoded in a representation. And maybe in machine learning, like all these attention mechanisms which have now been developed could be a manner to go through it. Maybe there are already attempts in these directions, I don't know. But this thing, at least among the things that I personally know, I would say that this is at least a partially, well, it's a difficult and least for what I know, unsolved problem. I don't know if you agree. So I'll come with a microphone. While we are at this discussion, I mean, Patrick and I, we were talking about this whole aspect of collaborations. And that's exactly what you guys were at. So collaborations are hard, like computer scientists speak different language or like machine learners speak different language, physicists speak a different language. And then whenever you would like to collaborate and then get fruitful interactions out of both of these communities, usually you would need some sort of a way to sort of translate one problem into the other and vice versa and see what tools one could use to solve the problem, right? So in your experience, working in your respective domains, what has been the best way to translate your problem? So let's say you have a physics problem, and then you'd like to use a machine learning model on it. What has been the best way for you to translate the problem for a machine learner? Because then usually they are thinking in terms of either text or images. And then usually there's a lot, a completely different structure. And then it's really hard for somebody who's not thinking in terms of the same things to translate their problem. So I think the question is more of like, how is your experience been with collaborations and what are the best practices that you guys have seen in your work? My experiences have been essentially very positive, but you should be aware that you need to take the time and energy to build the bridges. You really have to have the computer science people and the materials people sit and work together on meaningful problems and over a number of iterations arrive at the common language. You can't do this in a week or in two weeks, but honestly I view this whole event and the series of events as an effort in that direction to build this bridge between the machine learning and the materials people. And this should be really the tip of the iceberg, what you see at this event. But further down in the levels that you don't see that below the water you should have lots of joint ventures between machine learning and materials people brainstorming on how to solve these crucial issues. The materials people should be describing this to the computational people. What they think material is, you know, crystal structures, molecules, all of these aspects and then the computational people should be really working on formalizing these representations in a way that other computer science people would understand them and in that way opening the problems to the broader computational sciences forum. I'll make a short one about language which and grammars which I think is a recruiting team both in collaborations as well as in defining material or defining the good grammar like what I asked Francesca during her talk or the attempt by Johannes and Zachari on a grammar for catalysts. So it's becoming really a fundamental scientific and not only philosophical question related to languages and I'm really fascinated by it. A very quick response to that Kevin. We should be aware, you know, today data driven approaches are extremely popular to the extent that earlier knowledge driven approaches to solving problems are getting forgotten. You know, before learning grammars in a statistical way people were actually trying to construct grammars by hand and while obviously the data driven construction has its advantages, I think the real meaning is the combination of the two that you on one hand have let's say sketches of what the grammar would look like and now I'm talking really of grammars in different areas both for natural language and for molecules and whatever other problems might show up bringing together both the data and existing knowledge this is really the winning formula even though it might not be the easiest one. You know, I'm a machine learning person by education but broad enough to consider let's say more powerful formalisms for representing data and knowledge that come from logic and you know people are lazy they would if you tell them that besides the data they need to provide knowledge they will snub you off and say what you want me to provide knowledge in addition to the data but really you know it's a way to do things better and you know every investment you make there can pay off and the good thing is today you have no longer just databases of data but you have also in a way of knowledge you have ontologies you have you have databases of models in certain areas like in systems biology there is biomodels.net it's not only the data that is published it's also models that are published and once you get to that stage you know that you have repositories not just of data but also of knowledge then you really can work on this connection between data and knowledge and get to new levels but it does take effort you know it's easier to just use data it's just the easiest it's not always the best thank you yeah okay so thank you uh I I was thinking that because there is just a great abundance of different machine learning methods there like a dozen of well established neural networks Gaussian processes kernel method decision trees etc etc and we all have our favorites so would it be like what do you think is it better that one person more like tries to be an expert on just the one type of methods or focusing on one types of methods or would it be better to be like an expert a little bit everything like I try to be like this kind of a jack of all trades or would it be just better to focus on one type of method so I started actually with kernel methods and at some point I switched to neural networks but I think it still helps me that I know kernel kernel methods and were taught the basics there and something that everybody should understand is a linear model you should really there are some things that might be unexpected especially in high dimensions and this is something that I think everybody should look at and then there are other things like decision trees for example I never looked at but I think that is perhaps also a good thing to also then look in these areas because there's always some potential to combine methods and create something new so it's always good to have a broad overview view and then so basically specialize not too early start specializing once you have a bit have had a look around and then specialize but don't don't disregard everything else I would like to comment briefly so in the in the machine learning community there is nowadays lots of effort on something which is called auto ml you know finding so that you have automated systems that try different machine learning approaches for you and then give you back the best performing machine learning method but the prerequisite to use that is a very good understanding and formulation of the problem I mean I'm new to machine learning in materials but I have been trying to use just methods from my machine learning portfolio on problems and then I find that the current solutions that they use in material sciences are very you know matching kernel methods to predicting the energies of configurations of atoms at the at the kernel level without really formulating the problem as a predictive modeling problem one step back so you just you just go directly you know into the intestines of the problem very very deep and you don't even have the very general formulation problem that would allow you to try different machine learning methods on it so I've been looking at the publications and okay at that point I know how to use kernel methods on that but I don't know how to use other machine learning methods on that because the problem is not formulated in a general enough way so I think formulating the problem as you know there are categories of machine learning problems predictive modeling uh clustering approximating probability distributions these things so once you you specify things precisely enough then you could go for solutions like even auto ml if you if you really are just interested in the performance not not necessarily in the understanding if you want to understand the models that are coming back from machine learning then you need to worry a little bit about using explainable machine learning methods rather than just any black box but this depends on what your requirements are for solving the problem one question one new question yeah so I'm in machine learning or from a well-known researcher Richard Sutton there's something called the bitter lesson which is basically that the experience in the last decades of machine learning has been that in the end methods that leverage computation and data the best or even leverage the most computation have been the most successful in the end so we have seen this in natural language processing image recognition and so on and so the question is do you think this bitter lesson also applies to machine learning and material science that in the end the incremental advances through human intuition and more domain knowledge will be overtaken by just leveraging more computation so will we also have to learn this bitter lesson what do you think it's a philosophical point to some extent but I would say my religion is that I believe that knowledge counts and that even if you can arrive at predictive accuracies which are which are high by just using brute force more computation I still think there is an advantage to generating models which are understandable and can be used by people in addition to just the computers I would also like to add here as Andrea said very nicely in his talk data annotation is extremely difficult so there are a lot of challenges that are not solved yet and where we don't know how to actually get the data so it's not only about leveraging computational power but also about how can we actually get the data that we want to learn well I think a problem related well specific to material science material modeling is really transferability this was already mentioned before in the sense that we don't really want to simply to extrapolate our interpolate our prediction but we want to extrapolate our prediction this is a problem that doesn't really appear in natural language processing and this makes a difference in my opinion I think on okay one last question from zoom sorry would you like to comment no no I was screened for closing remarks so please I see so there was a question in zoom where do you see the potential of analog computing or machine learning methods a bit of a left field question analog computing is a tool for making things faster when they are very very specific so here the question could be actually reversed is there a specific machine learning application where the speed in the response is crucial and in this moment is low then in this case one can imagine to well to go in the direction of to go in the direction of analog computing which is a very very problem specific typically very difficult it requires a lot of specific knowledge it's not like software somehow no so here I wouldn't be able to tell an application in machine learning where this would be the case personally okay thank you in the interest of time I need to close this panel session but I hope you also enjoyed listening about grammars extrapolations and new problems we I mean old problems that we need to face thank you very much to everybody