 production on the development of new glasses and what are the current challenges of our domain. Glasses, unfortunately, are hampered by very inefficient trial and error design discovery methods. And it's because it lacks crucial access to consistent information. This because the glass production process is very much focused on details and nuance, but it's hard to report all of these details in a formal manner. And that's why we can't really build any models on the process parameters, but they are the most important information we got. And on top of that, we have lots of complex property relationships because of the disordered nature of our material. So if you're an experimentalist or material scientist working with machine learning, what are the sources of information that you can work with? So first of all, yes, of course, the composition. It might seem consistent, but it's really not because oftentimes you have quite a few volatiles in your glass composition, which is why you end up with a completely different product that you batched your oven with. And on top of that, there is a limited impact on your properties because, yeah, you can have a fully crystallized material and a glass with the same composition, but you have no idea how we ended up at one of either. So the second important information is, of course, the glass structure, but it's hard to access in a large scale. So if you're an experimentalist and you want to do thousands or tens of thousands of investigations on glass structure, you need to perform a lot of IR or Raman or NMR spectra. So this is not quite feasible. And there's, of course, a lot of inhomogeneous surfaces and bulk materials because sometimes you want to introduce some kind of stress within your glass or some kind of chemical gradient from the inside to the outside or the other way around. So it's, yeah, quite hard. The best source of information to focus on is, therefore, the process parameters, but we have a lack of sufficient data and it's expected to have a very high impact on properties. So our first approach was to try to get as much information out of the composition as possible. And we did a quite straightforward approach by using up-initial descriptors. There isn't really an Oxford dictionary definition or trademark on this word. It's just interchangeable with composition-weighted average of glass-constituent properties. So we did a bunch of DFT calculations for glass constituents, such as S02 or sodium oxide. And we then just built a property database out of that and weight the glass constituents accordingly to their more percentage-referring glass. So we have a rough first estimate. It's, again, very simple straightforward approach of our glass. What we did to verify the feasibility was to use a very common benchmark for glass science. It's the CyGlass dataset. You might not say, well, we have a large data set, but the CyGlass dataset is very dated. It had a sort of corrupted information within it and it's also not covering process parameters. But we could show that with this benchmark, we could achieve similar results with our up-initial descriptors than we have with the compositional descriptors. From another experimentalist group, we also got a very small-scale database because right now, since we have no comprehensive large-scale data sets, we tried to fulfill the needs of these people. So we got from the Wondracek group as a trend-rate sensitivity database that covers only 340 entries with 57 unique components. And it covers not only oxide glasses, but also a lot of metallic glasses and child-cooking need glasses. And we could show that we got at least a bit of a performance increase throughout our models for the exception for the neural networks because we believe that only seven up-initial descriptors and only 340 entries is not flexible enough for this type of machine learning model. But since this is a very inexpensive method, once you obtain all your DFT calculations, you can always give it a shot as an experimentalist. And what we found out doing our process was that we could enable expandability without retraining our models. Because in this up-initial descriptor space, we have one common input space for all types of glass compositions. And we're now exploring what kind of applications we can explore with that. Moving on to the second source of information, the structural descriptors. As I said, usually you get access to the glass structure as an experimentalist through IR, NMR, or Raman spectrum. But this is not automated. At least it's not automated yet. We have an experimentalist group as our partners that are working on building some fancy robot facilities to do that. But as of right now, we can't really work with that. So we need simulations once again. And to build upon existing information is also quite difficult because usually people are interested in some very specific systems, some very specific properties. And you can't really build a very dense input space out of that, obviously. So our idea was why not remove our bias from the equation and let a machine decide to interpret what defines a glass structure. And yeah, but first I want to talk about the simulations, of course. We want to use MD, or we are using MD simulations to build next to the experimental data pipeline. We want to build a simulation data pipeline that is continually speeding our database with information. So we're randomly generating structures according to the glass composition. And yeah, depending on the target property, we are starting with about 50,000 simulations. The cells as of right now cannot be too large and the simulation protocol cannot be too extensive because we have only so much time, right? And yeah, throughout this conference, there were a lot of fancy force field methods presented, but since we wanted to keep it simple and straight at the beginning, and we are using a very common force field within the field of glass science, the force field according to Perdon et al. But later on, we can fancy up on that part. All right, so our idea to extract the structure from our, or to extract the information from our structures are better variational autoencoders. There have been quite a few talks about variational autoencoders, so I don't want to spend too much time on what they are or what they are doing. The key difference is this, yeah, better. So we want to ensure that the latent space is as disentangled as possible, so we have as non-correlated latent factors as possible. This, of course, is on the expense of the reconstruction error, so we probably won't have as good of a reconstruction as without the better. But by talking to experimentalists, this does not need to be an disadvantage for glasses since to an extent, we do want to have some variability in the reconstruction of our disorder structures. So since throughout our data pipelines from the simulations and the experiments, we expect to have a very dense input space. We expect to have continuous latent factors, which are very exciting in our opinion because we can then attribute these latent factors to some kind of properties. As of right now, since we are in the process of doing this, we don't know if these latent factors are physical or non-physical. But either way, we're looking forward to have some kind of unbiased insight into the glass structure and find some way to create maybe some generative model or some preview throughout our glass systems. And I'm not sure how much time I've left or how much time I've spent. So I want to thank my supervisor, Marek Zürka and, of course, the German fellow government for funding this project and also all the project partners from Claustar, Würzburg, Jena, and Berlin, especially Berlin, who are building these awesome robotic facilities. And yeah, I want to let conclude my talk. Thank you, Felix. How about questions for Felix? Yes, Antti. Thank you for your presentation. I have one quick question about this. So glasses are amorphous materials. So what kind of challenges this introduces compared to, for example, these crystalline materials for applying these your methods? Yeah, well, in the field of crystallized materials, you often can work with symmetries or representation of symmetries, as we have seen throughout this conference. And for glasses, to an extent, we have some short-range order or medium-range order, but there's no long-range order, and we can't really use methods of symmetry to that extent. So that's why there's a need of alternative methods and also for the inclusion of process parameters, because we need to somewhat work on a larger scale than lots of the molecular guys. I have a question here. Could you just go back to your slide with your autoencoder diagram? So do I understand correctly you want to encode and decode the entire unit cell, not just individual environments at a time? Yes, the entire cell. OK, gotcha. And does it work? Or are you not there yet? Well, we're not there yet. We have trouble to set up this simulation pipeline in a way that we can scale it easily, and we have a reliable output. And yeah, so we're not there yet, but this is the idea. Yeah, very exciting. You like? OK, any further questions? Yes. What are the tricks you would use to account for periodicity in the unit cell? What are the tricks needed? Or is it natural to account for the periodicity of the unit cell? We lack periodicity. We don't have periodic structures, like we do, but we have these other structures. But in my group, some other people show that for polymer materials, like at a later point we could do more calculations for the same glass and thereby generate larger cells, because we can just stitch them together and we have some inherent variability because of the randomly generated structures. So I think maybe we'll ask that your question even. Yeah. You might be out encoding a finite size effect if you don't. Yeah, we want to encode more of the same compositions, but also more other compositions, obviously. Yeah, thank you. OK, that seems like I can ask my question, which is about variational OTIN courses. And I'm very interested in latent spaces. Have you looked inside your latent space and have you looked into where your structures are in latent space or any kind of just qualitative analysis? In our latent space? Or not yet? Yes, very much. OK, that's maybe in the next meeting. Yeah, for sure. Any last questions for Felix? If not, let's give him a hand and all the speakers of our session today. Thank you very much. And next, I think we have our panel discussion. So am I right? This is the.