Podcast: Crawl, Walk, Run: AI Transforms Materials Discovery

Transcript

Edited for clarity

Traci: Welcome to Chemical Processing's Distilled Podcast. This podcast and its transcript can be found at chemicalprocessing.com. You can also download this podcast on your favorite player. I'm Traci Purdum, editor-in-chief of CP, and today we're discussing the scientific frontier of materials discovery and what's next in scientific infrastructure and AI for science.

To help me navigate this frontier is Joshua Young, senior application scientist at the U.S. branch of Japanese AI startup Matlantis, which uses atomic-scale, AI-based simulation to accelerate global materials discovery, specializing in computational materials research. Matlantis receives backing from Preferred Networks, ENEOS Corp. and Mitsubishi Corp.

Joshua points out that in the chemical industry, AI and machine learning are allowing chemists to simulate molecular systems millions of times faster than traditional methods, screen thousands of candidates at once and directly inform product development pipelines. There's a lot for us to cover today. Joshua, thank you for joining me. Would you share a bit more about your background?

Joshua: Thanks so much for having me. My background is in computational materials and computational chemistry. I have a bachelor's in chemistry and a Ph.D. in material science, and I've been involved in the material simulation industry and research for about 15 years now.

Traci: In those 15 years, a lot has changed. Everything is moving so fast lately, so our conversation today will hopefully get us up to speed on what's happening. AI seems to be the buzzword on everybody's bingo card, and in the chemical industry there are a lot of challenges. Can you go over a few of those for us?

AI Challenges

Joshua: There are a lot of challenges involving AI in the chemical industry. AI tools are powerful, but we're seeing some roadblocks in their successful implementation — though we're also seeing a lot of successes. We see the primary challenges of AI adoption in the chemical and materials industries as threefold.

First, there's what we call the data problem. Many AI projects are hindered by small data and messy data. In the chemical and materials industries, it can take thousands of dollars and work hours to get a single data point. Then you have all this information generated and scattered across lab notebooks, spreadsheets, microscopy images, characterization spectra and kinetic data. It's unstructured, noisy and hard to train reliable AI models from scratch. They need well-structured data and lots of it.

There's also what we call the black box problem. As chemical engineers, we're usually trained to rely on first-principle models — we understand chemistry and physics, and we can write down equations that explain what we observe. Often now we run into AI models that can't explain why they reached a conclusion in a way that aligns with physical laws. For an industry built on safety, this lack of interpretability can be a significant barrier to trust.

The third challenge is the skills gap problem. There aren't enough trained chemical engineers or material scientists well-versed in the applications of machine learning and AI to engineering problems — not only at the R&D level but also across all types of teams. This makes it hard for companies to integrate a cohesive AI plan. You need employees and management capable of identifying the specific areas where AI can make a high-impact difference.

But despite all this, there are a growing number of successes. AI has been found to solve a lot of problems in process optimization, predictive maintenance, digital twins of manufacturing processes and automated experiments. These are very exciting and gaining a lot of traction. From our point of view, one of the most profound successes is happening in accelerating materials research and discovery — and chemical engineering is full of problems that are fundamentally materials problems.

You mentioned that one of Matlantis' key stakeholders is ENEOS, so we have a long history in the chemical industry. Catalyst development, reaction engineering, polymer synthesis, membranes and separations — these all have materials challenges that need to be solved. The traditional R&D cycle is limited by the speed of trial-and-error experiments and even conventional simulations.

Our primary offering, is an AI-trained atomic potential trained on a massive set of materials calculations. By applying this model, it's possible to run simulations to predict material properties for reaction processes that once took days on a supercomputer in just seconds on a laptop. Our clients have applied our AI-based software and platform to do materials and chemical prediction and development, accelerating the R&D process. So yes, there are a lot of challenges in implementing AI, but a lot of success stories as well.

How Exactly Does AI Speed Up Processes?

Traci: In Chemical Processing, we've been writing about AI and some of those success stories, mirroring exactly what you've said. We've seen that folks at 3M and DuPont are lowering their timeframes from years to months, and that's incredible. You've set the stage and given us a lot to think about. I want to get more granular on some of these things to help our readers understand how they can utilize AI for their benefit and make sure they're not afraid of this technology.

Let's get specific: When a chemical engineer is staring at a corrosion problem or trying to optimize a catalyst system, how exactly does AI speed this up?

Joshua: That's a good question, and there are a few different ways. In my opinion, they fall generally into two categories: You can apply AI to learn from past experimental data to accelerate future R&D, or you can accelerate future discovery through simulation.

First, on the experimental side, AI can analyze historical data from lab notebooks or databases to find hidden correlations. Say you're investigating a corrosion problem and you digitize everything, get it ready for an AI tool and feed it in. The AI model might discover patterns that are imperceptible to humans — a human might miss them. That's what AI models and machine learning are really good at: pattern recognition. It might discover that some specific trace element, when you combine it with certain processing temperatures, consistently led to lower corrosion rates. This is insight you can gain from adopting AI tools or machine learning models that find relationships in your experimental data.

The other way is accelerating future discovery through simulation. Let's say you want to design a new catalyst, and you have a hypothesis that if you take your catalyst and dope some new compound into it, it's going to lower the energy of some key reaction step. The very nice thing about simulation is that you can control a lot of things to study their effect, which would be impossible to control or investigate in experiments. That's why we need both simulation and experiments in the chemical and materials industries.

Traditionally, though, simulations have been rather slow. If you have a trained AI model like the one provided in our platform, you can quickly investigate hundreds or thousands of different catalyst compositions, look at all the effects on the key reaction step for whatever reaction is of interest to you and say, "OK, my hypothesis was shown to be correct or not correct," and if it was correct, "This is the specific chemistry that would give me the best performance." You take that back to your experimental colleagues, do some more experiments and then close the loop. You either turn that into a product or you come back to the simulation.

Really, what AI has been able to do is speed up the simulation side. Our platform can perform these calculations now in seconds that used to take days or weeks with conventional simulation methods.

The Messy Data Problem

Traci: Obviously, utilizing AI this way, it's a companion. It's not taking over. It relies on real people to take it that step further. So it's doing a lot of the work that, as you pointed out earlier, involves the data these folks have sitting all over the place — their legacy systems, Excel spreadsheets. Our industry loves Excel spreadsheets and lab notebooks. How do you tackle this messy data problem and get it to make sense? You alluded to having this big machine learning catalog to go through everything. How do you pull that all together?

Joshua: That's an excellent question, and I think that is one of the biggest barriers that kills most industrial AI projects before they even get started. It's a hard question, and I don't know if I have 100% of an answer, but it depends on what exactly you're trying to do.

What kills a lot of these projects is companies trying to jump in too fast. You come in and say, "OK, we've got Excel spreadsheets and microscopy images and all sorts of data stored in different formats, so let's bring in all these AI tools, digitize everything, and now we'll be efficient and discover all sorts of new things quickly." That will actually end up interrupting workflows. They have to learn all sorts of new tools, and it's still a question of how you deal with all these different formats of data.

I think the best way to start is to pick one high-value problem that you think AI can solve. It could be optimizing specific parameters to maximize yield or something like that. You start and curate just enough of your data to build small machine learning models, small AI models, and prove that you can get a tangible return on investment from implementing some AI tools and investing in what's actually a data management problem, not really an R&D problem. Once you have that initial success, you can justify more investment into other digitization tools, moving things over to digital lab notebooks and things like that.

Another option some companies take is to start generating new data from scratch. Just start a whole new project, begin running your experiments and set up the data management plan beforehand. Set up automated, high-throughput labs to create large, clean datasets. This works, but you lose the accumulated knowledge that companies with decades of experience have, and it can be a slow and expensive path.

This is actually the approach we took. Matlantis and similar simulation platforms have used piecemeal datasets, open-source datasets and things like that, trying to piece data from different sources together to make a training set. What we did was start our own dataset from scratch. This was about four years ago, and we said, "We want to create a training set that allows us to be the most expansive in what we can cover in terms of material structure and chemistry." I believe we've successfully done that. We now have more than 60 million quantum chemical calculations that go into training our AI model, which is then hosted in our platform.

Those are the two approaches I can see, but there's not really, unfortunately, a one-size-fits-all solution. It depends on the project, it depends on the data you have, but I think: Take it slow, make a plan.

Winning Over Skeptics

Traci: Getting in there, getting used to it and understanding its capabilities — I think that's key for any new technology. You mentioned it earlier too: Our industry is very conservative. They're skeptical. Engineers see all these game-changing technologies all the time, and unless they're absolutely sure, they just don't want that buy-in. Can we talk a little bit about that? How do you build confidence that this type of tool will help them in their work? And you talked about open source — what about the security aspect? What about the safety aspect of everything?

Joshua: As engineers, you're correct — we are skeptical. We always have to think about the safety aspect of things, and we have a deep respect for the skepticism. Our team is full of chemical engineers, materials engineers, chemists, people from all sorts of these fields, so we have a deep respect for it. It's essential for good engineering.

In terms of validating AI models and getting buy-in, there are a couple of things. First, the accuracy and performance of your AI model is very important to building trust, regardless of how you measure or evaluate that. But that's not the whole story, because you can have some perfect AI model predicting what you want to predict exactly with very low error and things like that. But how do you actually trust it and implement it into workflows?

In my opinion, one of the key ingredients to getting this type of buy-in is time. Some companies are more risk-averse, some are more risk-tolerant, but even the most risk-tolerant companies won't suddenly implement AI models at the top level without some kind of trust. You build trust through a validation framework, and one of these is what I'll call the "crawl, walk and run" approach.

Let's say you've been hearing about all these AI tools and want to implement them in your workflow. You don't just start out and push the AI into production. As you mentioned earlier, we view these AI tools as tools that assist human researchers in making decisions. What we encourage our customers or potential customers to do is first do a low-stakes trial.

After that, some trust is generated. Maybe you compare it against existing results, existing experiments, existing simulations — something where if something goes wrong, it's not going to be disastrous. So we're crawling. We go through that, and the users are satisfied.

Then we move on to walking. You adopt Matlantis — and this could go for any AI tool — and now the engineers are using it on new problems. However, they're comparing it against their own thoughts, decisions, processes and workflows. Maybe they're applying it to help develop some new material, actually involving it in a project.

Then finally, after the AI model has proven itself in both the crawl and walk phases, you can gradually allow it to run and become more integrated into workflows. Maybe now the AI model, the simulation results and the engineer disagree, but now that there's that trust, maybe the engineer will think, "OK, the AI might be onto something," double-check it and find out it actually is correct. Or if you think about other types of AI models, maybe you start letting it actually give some kind of advisory alerts, or you give it some automated role.

That's the validation framework that I feel allows engineers to gain trust in these game-changing technologies. And some companies might not even go from the walk to the run stage — they want it to just be another tool in a toolbox. And that's perfectly fine. There's no one-size-fits-all solution.

AI in Five to 10 Years

Traci: Absolutely. You can use it how you want to use it, and I think you'll eventually get to the walk and run after you realize how useful it can be. Speaking to some of our readers, they don't even understand how it can be applied to them. It's just, as I said before, that buzzword that they're not even certain how it would apply to their work, and we're learning more and more how it can do that. Conversations with you help as well.

I want you to put on your Nostradamus hat and look into the future a little bit for us. Can you tell us anything about what you see coming in five to 10 years — how AI is going to change the workflows, materials discovery and chemical process development? You mentioned before in silico. Let's talk about that.

Joshua: Yeah, absolutely. Even thinking five to 10 years from now, I would argue that a lot of this R&D workflow is already being imported into an in silico-first model. In five to 10 years, I think much of materials development and chemical development will move to this in silico-first model. Because of the nature of classical simulations that we're used to, I think today's workflow is pretty experiment-heavy and simulation-light. We see a future where it's simulation-heavy and experiment-light, saving time and money on expensive experiments by going to fast, AI-guided materials or chemical simulations first. This is actually what a lot of our customers are doing right now.

The new workflow will look something like this: First, you go into discovery mode. You screen thousands, hundreds of thousands, millions of hypothetical materials using tools like Matlantis, exploring all possibilities that would be impossible to test physically or experimentally. Then you go to development. You choose promising candidates, you can do more complex simulations under realistic processing conditions, optimize performance also in silico. And then finally, you go to validation with the physical lab work — beakers, flasks, batch reactors — to validate your promising candidates that have already been extensively vetted in silico.

We see this leading to what we'll call the "autonomous discovery loop," where AI tools like Matlantis help you design novel materials and see their performance — like we discussed with the catalyst discovery process. You want to optimize, lower the energy of some key reaction step. Then you can pair that with other AI tools: Automated experiments and laboratory robotics are getting extremely popular to do high-throughput composition optimization. You can check the whole gradient across some alloy, some metallic alloy catalyst or something like that, and test them. You feed that back into some kind of machine learning model, gain more insights, learn more patterns and keep going in this loop to accelerate the R&D process.

Traci: As with anything, there have to be some concerns and challenges and things to look out for.

Joshua: The key to this is validation of the tools, and that's why I mentioned this crawl, walk, run approach. I think implementing tools blindly is not the best recipe for success. You can run into all sorts of problems trying to do that. Take the time, see what the tools are doing, what they can offer you, try them out, make sure you understand them and then implement them. That's more from a management side of things than the R&D side. Identify key bottlenecks, try to apply tools, see which tools can help you overcome these bottlenecks and then introduce them.

Traci: Joshua, was there anything you want to add that maybe I didn't touch on?

Joshua: I'll just add that it's a very exciting time to be involved in material simulation and AI. There are all sorts of breakthroughs happening, and to me, it's really rewarding to see our customers have these types of R&D success stories by applying AI and simulation — what I spent my Ph.D. and postdocs doing. It's very exciting to see all of that come to fruition nowadays.

Traci: Exciting. Joshua, thank you for helping us with the transition and the crawl, walk and run approach. I like that, and thank you for breaking everything down.

Listeners, if you want to stay on top of innovation in the chemical industry, subscribe to this free podcast via your favorite podcast platform to learn best practices and gain insight. You can also visit us at chemicalprocessing.com for more tools and resources aimed at helping you achieve success.

On behalf of Joshua, I'm Traci, and this is Chemical Processing's Distilled podcast. Thanks for listening. Thanks again, Joshua.

Joshua: Thank you so much, Traci.