Schrödinger advances materials informatics for faster development of next-gen composites

Materials informatics (MI) is the basis for a digital transformation of the materials industry, using data infrastructure, machine learning (ML) and AI solutions to develop new materials and optimize how they are processed. According to IDTechEx — which has been tracking this technology for years — MI is a paradigm shift, dramatically reducing time to market. But it goes even further, not only accelerating the forward direction of innovation — realizing properties for suggested candidate materials — but also enabling inverse development, where novel materials are suggested based on input of desired properties.

An example of how MI is being used in composites was announced by Toray Industries (Tokyo, Japan) in 2021 — it leveraged MI to develop a new high-performance carbon fiber-reinforced polymer (CFRP) prepreg with exceptional fire-retardant (FR) performance for advanced aerospace applications. CW will report more on this soon.

“It’s a totally different approach,” says Anand Chandrasekaran, product manager for Material Science Informatics at Schrödinger (New York, N.Y., U.S.), a company that collaborates with customers to use its predictive computational tools to develop future materials, including composites. “Not only are we making material performance predictions many orders of magnitude faster than with traditional experiments, we are enabling a different type of ideation. You can now start with 10,000 or 100,000 possible structures or look at all the possible molecules you could purchase and put those different sets into a database which you then use ML to evaluate and downselect much more efficiently for the actual materials to synthesize and test.”

But Chandrasekaran adds that Schrödinger’s platforms also combine ML and physics-based approaches, enabling digital generation and validation of datasets. And this is opening a new universe of possibilities where existing datasets to train ML models may be thin, including in biomaterials, fire resistance and high-temperature composites, new materials for space applications, cryogenic hydrogen storage and much more.

Enabling more sustainable, high-performance composites

Before delving into the details of how MI works, it’s helpful to contextualize how and why it’s used. “R&D scientists across broad industries face challenges in developing the next-generation of polymers and composites that are high-performance, multifunctional and meet society’s demands for sustainability,” explains Schrödinger’s website. The company’s digital platform performs simulations of bio-based polymers and additives at molecular and atomic scales, enabling scientists to understand and predict performance. It can also automate screening of bio-based polymers, predicting structure, miscibility (ability to mix and form a homogeneous solution) and properties of polymer mixtures. It can also model water uptake and predict glass transition temperature (T_g), thermal stability, thermal expansion and polymer gel point during cure for new resin formulations targeting high-performance applications.

“There are many lifecycle questions being asked now,” says Andrea Browning, director for Polymers at Schrödinger. She works with Chandrasekaran and his colleagues on the physics-based tools side to offer a collection of capabilities for polymer and composites customers. “These customers are looking at different input monomers than were considered in the past. We enable leveraging their past data to understand possible connections to new monomers — including from different sources that they weren't considering before. MI enables you to leverage beyond what you might do in a traditional experimental reformulation study, so that you’re using it as a discovery tool.”

“Improved additives for flammability resistance, plasticization or some other target is another area where MI provides new possibilities in improvements,” says Browning. “This could include looking at new additives that are better for the overall life cycle of the material. And when exploring what other materials you might use, you again leverage the datasets that you might have, and then connect the characteristics of these new materials back to the chemistry. This gives you additional ways of looking at how you might combine and use new chemistries to meet both performance and sustainability targets.”

Exploring next-gen space materials

An example of how MI is being used for composites is at the U.S. Air Force Research Laboratory (AFRL) Aerospace Systems Directorate at Edwards Air Force Base, where research chemist Dr. Levi Moore and his team have published on understanding and exploration novel formulations for polycyanurate resins using the Schrödinger platform. Comprising cyanate ester monomers, these thermosetting polymers offer exceptional FR properties, important for high-temperature (HT) applications in aerospace, with multiple sources citing a Tg of ~300°C and char yields >70%, the latter interesting for ceramic matrix composites (CMC). Alternatively, they may including silicon (Si) for additional HT and char yield benefits.

Moore’s team wanted to explore polycyanurate reaction chemistry, but this is challenging because the curing/crosslinking is often complicated and requires considering multiple factors simultaneously. Specifically, Moore wanted to better understand water uptake. Even though this is low for polycyanurates, where it does occur, failure can result if that water experiences a rapid rise in temperature.

His team used Schrödinger’s computational platform to tackle these challenges. As Moore explains in a previous article, “We’ve been able to watch water diffuse into a crosslinked polymer matrix and see where the water molecules are most attracted to specific chemical groups. To do something like this experimentally, you would need access to extremely complicated setups only available at national labs, and even then, you would only be able to see the water on the macro-scale without the fine atomistic detail the simulation affords.” He adds that this digital simulation reduces “the number of new molecules we’d need to synthesize, leading to a more efficient and less resource-intensive iteration cycle.”

Polycyanurates are just one of several areas where Moore and his colleagues have applied Schrödinger’s MI tools. Researchers at AFRL are also exploring the potential for additive manufacturing and 3D printing to reconfigure existing materials in new, more complex geometries that may increase their tolerance to harsh environments. With publications, webinars and interviews, AFRL is providing an example of how aerospace companies can add MI to their research cycle.

Moore’s previous interview highlights a direction for future materials modeling: “The greatest challenge will be the end goal of a system where you can input the desired properties of a material and the system would be able to find a solution in that space, and if that solution does not exist, then it can iterate through new, as-of-yet undiscovered materials, and output a target synthetic molecule that will possess the desired properties.”

How does MI work?

One way, says Chandrasekaran, “is to use ML with existing datasets to make predictions about potentially new materials or to find underlying mechanisms to improve the properties of existing materials. And these ML techniques are fairly recent, developed in the past 8 years or so. At Schrödinger, we are developing ML technologies that take a particular material — it could be a polymer, a polymer plus an additive, a molecule or a mixture of molecules with a solvent, for example — and break that complex material down into its individual components and structures. We then have special featurization techniques to represent each structure in a way that can be digested by a computer algorithm. Once that material is digested by the algorithm, we can then find correlations between the structure and the composition of the material and the property of interest. We can predict performance and see how properties change.”

“However, when you have really small datasets, these ML predictions can be pretty off,” notes Chandrasekaran. “This is because ML is based on looking at statistical correlations from past data to make future predictions. But once we combine ML with physics-based approaches then we can get really powerful descriptors for the material on a first principles level — which means you don't have to make any assumptions on past data or on the statistical models. Instead, from the basics of quantum mechanics and molecular dynamics, we can generate the properties of materials. And even if you can’t generate the exact same properties of the composite that is required, there are many properties that are very good descriptors or have a very strong correlation to the final property that could be of interest for a composites manufacturer.”

“So, by using the combination of ML and these physics-based approaches, we can generate our own datasets and see which of these properties are correlated with actual experimental values,” he explains. “What we then have is a powerful methodology that combines a first principles approach with a statistical approach that enables researchers to devise completely new materials, new additives and new composites.”

SABIC speeds polymer development

An example of this combined ML and physics-based approach is Schrödinger’s recent work with SABIC (Riyadh, Saudi Arabia, and Houston, Texas, U.S.). “We used ML on their in-house dataset of T_g, CTE [coefficient of theral expansion] and dielectric properties to come up with new monomers and copolymers,” says Chandrasekaran. “We made predictions on tens of thousands of new potential repeat units and polymorphic units, and then took the top performing ones as ranked by ML. We then validated those with physics-based simulations, and finally, the company was able to synthesize the most promising polymers.”

Chandrasekaran and Browning point out that the classic approach to designing new polymers relies heavily on trial-and-error experiments that are slow and expensive. Adding to this complexity, when researchers improve one aspect of a polymer, it often worsens other aspects. Thus, designing new polymers is like solving a complicated puzzle which requires optimizing multiple properties simultaneously. Instead, the resin design and incubation team at SABIC Specialties —  which supplies polymers, compounds and advanced materials into a wide range of composites applications — worked with Schrödinger’s material science team to build accurate ML models to speed up development.

SABIC curated a physical dataset of five target properties related to how polymers respond to mechanical stress, temperature, electrical and optical conditions. The team then successfully trained and validated ML models that accurately predicted these five polymer properties using only molecular structure as input. Schrödinger and SABIC then used these ML models to explore new polymer structures that had not yet been tested experimentally but exhibited polymer properties useful to real-world applications.

The team generated more than 10,000 structures, deployed the trained ML models to predict the polymer properties and applied a multi-parameter optimization (MPO) ranking criteria to identify polymer structures with preferred characteristics. Using Schrödinger’s LiveDesign MI platform, the team narrowed the design space to 1,000 structures and then pinpointed and downselected commercially available structures to 10 with promising properties.

To give further confidence in these 10 candidate structures, SABIC and Schrödinger tested them using physics-based approaches. The outcomes were consistent with those predicted by the ML models. SABIC then synthesized and physically tested the three most commercially viable of the 10 candidate structures. All of them satisfied all of its initial criteria for higher performance materials targeting applications in mobility, infrastructure, 5G, medical devices, coatings and more.

SABIC noted this approach can shorten polymer innovation timelines from several years to several months. Even further, this AI-driven workflow has also inspired scientists to leverage computational tools to test much more novel structures that wouldn’t have been feasible through trial-and-error experimentation alone. (Read more: “Fast-Tracking Next-Gen Polymers: How SABIC is Leveraging Machine Learning and Physics”).

A transformative approach

How significant is this reduced timeline? “When we narrowed down the 10,000 polymers structures to the top 10 polymers for physics-based simulations, each of those simulations might take a day or so,” says Chandrasekaran. “But to come up with a polymer experimentally or directly synthesize and test it would take at least a month or more. So, it's multiple orders of magnitude faster than just experiments, and this is truly transformative.”

“But you wouldn't even be able to do that if you just had access to experiments,” adds Browning. “You would bring together your experts in that area, and they would decide to try five or 20 or 30 — whatever you had budget and time for. You wouldn't be screening across numbers that are double orders of magnitude larger and really accessing this much broader realm of ideas. Instead, you would have to do a lot of narrowing down beforehand.”

“And it would take time to do that,” she continues, “and then once you narrowed it down, it would take time to source the constituents, synthesize the polymers and — if you’re doing something really new — there may be a couple of iterations just to make sure you made the polymer you intended to. Instead, we’re doing it all digitally, so that you get a subset with a higher probability of success, and you’re going into experiments with the ones you’re most confident in and verifying those. That cycle is much, much shorter.”

Chandrasekaran notes this also opens up new innovations not previously possible. “Say you have a really challenging task, such as replacing an ingredient because of environmental regulations with something that gives you the same properties or better but is no longer problematic regarding toxicity or sustainability. Using this MI approach, you can actually generate a huge library of compounds that are viable, and that can be bought and approved. You can then run ML and make predictions as to which of those ingredients and additives are most likely to help achieve the same or better properties required.”

“So, you can clear the plate and consider all the possible molecules that you can test and polymers that you can use,” says Chandrasekaran. “And you can easily list 100,000 compounds and then make predictions using ML in minutes. So, it's a totally different approach that allows you to start ideating in a different way. You can draw monomers and look at all the existing sources for the molecules that are available for you to purchase and then put all that into a database, using ML to scan through that database much more efficiently for the additives or molecules of interest.”

Possibilities for H₂ tanks and more

Another area where MI could have significant impacts is in current research into polymers that can resist cryogenic temperatures and prevent liquid hydrogen (LH₂) permeating through composite laminates for new storage solutions in zero-emission aircraft and linerless tanks for spacecraft applications. It seems like MI could unlock new enabling materials without years of experimentation. “Exactly, that’s the goal,” says Chandrasekaran. “And Schrödinger already has a product called Formulations Machine Learning, where any scientist can do ML on the desktop or laptop and tweak the component, the composition and the chemistry, and get very accurate performance predictions. You can make predictions on this whole database and then do experiments just on the best performing molecules.”

This and Schrödinger’s other MI tools and platforms are desktop/laptop-based software with different licenses based on their format and use. “And all the data stays within the company’s system because nothing goes to the cloud,” notes Chandrasekaran. “All the data the company uses and all their IP and important formulations — it all stays within the desktop or laptop and all the analysis and ML can be done locally on the user's machine.”

Browning notes the very wide range of applications for these tools. “We’re talking about polymers and polymer composites right now, but what you can do with this approach is much broader. Basically, any place that you're trying to design a material to do something under some kind of constraints, that's where these kinds of tools are very powerful. It's about having an understanding of what you need the tools to do. The hydrogen tanks are a good example, where you need to have strength and low permeability, as well as CTE performance during temperature cycling. So, having an idea of what your objective is and the key inputs enables these tools to be extremely useful.”

MI has also come a long way from where it was even 10 years ago, adds Browning, including the associated techniques around it. “For example, we talked a little about using physics-based simulations as a way to augment what you can do with MI, and those have come a long way. People who may have looked at this 10 years ago and said it wasn’t for them — I would say that it's very different now. Both the ease of use and the results that are possible have really advanced.”

“And we don't just let customers download the software and run it by themselves,” says Chandrasekaran. “We really want to work with them to make sure that this tool is valuable for them. So, we help the customer format the data and frame the question for their case and work with them to get the best value from these tools.”

Stay tuned as CW reports more on MI and how it is changing the development of new composites.