AI Foundation Models Drive Breakthroughs in Battery Materials

With entry to ALCF’s highly effective Aurora and Polaris programs, researchers are creating AI fashions that may predict promising new supplies for battery electrolytes and electrodes.

For many years, the seek for higher battery supplies has largely been a means of trial and error.

“For a lot of the historical past of battery supplies discovery, it’s actually been instinct that has led to new innovations,” mentioned Venkat Viswanathan, an affiliate professor on the College of Michigan. ​“A lot of the supplies we use right this moment have been found in a comparatively quick window between 1975 and 1985. We’re nonetheless primarily counting on that very same set of supplies, with some small, incremental tweaks to enhance battery efficiency.”

It’s like each graduate scholar will get to talk with a prime electrolyte scientist each day. You’ve got that functionality proper at your fingertips and it unlocks a complete new degree of exploration.” 

Venkat Viswanathan, affiliate professor on the College of Michigan

At this time, advances in synthetic intelligence (AI), and the computing energy to assist them, are altering the sport. With entry to supercomputers on the U.S. Division of Power’s (DOE) Argonne Nationwide Laboratory, Viswanathan and his collaborators are creating AI basis fashions to hurry up the invention of latest battery supplies for functions starting from private electronics to medical gadgets.

Basis fashions are giant AI programs skilled on huge datasets to study particular domains. Not like general-purpose giant language fashions (LLMs) akin to ChatGPT, scientific basis fashions are tailor-made for specialised fields like drug discovery or neuroscience, enabling researchers to generate extra exact and dependable predictions.

“The fantastic thing about our basis mannequin is that it has constructed a broad understanding of the molecular universe, which makes it rather more environment friendly when tackling particular duties like predicting properties,” Viswanathan mentioned. ​“We are able to predict issues like conductivity, which tells you how briskly you possibly can cost the battery. We are able to additionally predict melting level, boiling level, flammability and all types of different properties which are helpful for battery design.”

AI Helps Researchers Discover the Huge Chemical House

The group’s fashions are targeted on figuring out supplies for 2 key battery parts: electrolytes, which carry electrical cost, and electrodes, which retailer and launch vitality. Advances in each are wanted to design extra highly effective, longer-lasting and safer next-generation batteries.

The problem is the size of the chemical area for potential battery supplies. Scientists estimate there might be 10⁶⁰ doable molecular compounds. A basis mannequin skilled on knowledge from billions of recognized molecules may also help researchers discover this area extra effectively. By studying patterns that may predict the properties of latest, untested molecules, the mannequin can zero in on high-potential candidates.

In 2024, Viswanathan’s group, together with Ph.D. college students Anoushka Bhutani and Alexius Waddle, used the Polaris supercomputer on the Argonne Management Computing Facility (ALCF) to coach one of many largest chemical basis fashions up to now. The mannequin is targeted on small molecules which are key to designing battery electrolytes. The ALCF is a DOE Workplace of Science consumer facility that’s obtainable to researchers from internationally.

To show the mannequin find out how to perceive molecular constructions, the group employed SMILES, a extensively used system that gives text-based representations of molecules. In addition they developed a brand new software known as SMIRK to enhance how the mannequin processes these constructions, enabling it to be taught from billions of molecules with better precision and consistency.

Constructing on this success, the researchers at the moment are utilizing the ALCF’s new Aurora exascale system to develop a second basis mannequin for molecular crystals, which function the constructing blocks of battery electrodes.

As soon as skilled, the inspiration fashions are validated by evaluating their predictions with experimental knowledge to make sure accuracy. This step is important for constructing confidence within the mannequin’s capability to foretell a variety of chemical and bodily properties.

Previous to creating the inspiration mannequin, Viswanathan’s group had been creating smaller, separate AI fashions for every property of curiosity. The inspiration mannequin skilled on Polaris not solely unified these capabilities beneath one roof, it additionally outperformed the single-property prediction fashions they created over the previous few years.

The group is actively exploring the mannequin’s capabilities and intends to make it obtainable to the broader analysis group sooner or later. The group additionally plans to collaborate with laboratory scientists on the College of Michigan to synthesize and take a look at probably the most promising candidates recognized by the AI fashions.

Scaling Up with Argonne Supercomputers

Coaching a basis mannequin on knowledge from billions of molecules requires computing energy that’s past the in-house capabilities of most analysis labs.

Earlier than having access to ALCF supercomputers by way of DOE’s Progressive and Novel Computational Impression on Principle and Experiment (INCITE) program, the group was working into scaling points. Bharath Ramsundar, a part of the INCITE challenge group, had constructed AI fashions skilled on tens of hundreds of thousands of molecules however discovered they may not match the efficiency of present state-of-the-art AI fashions.

“There have been sharp limitations within the variety of molecules we might take a look at when coaching these AI programs,” mentioned Ramsundar, founder and CEO of Deep Forest Sciences, a startup firm specializing in AI-driven scientific discovery. ​“We began with fashions skilled on just one million to 10 million molecules. Finally, we reached 100 million, however it nonetheless wasn’t sufficient.”

The corporate has explored utilizing public cloud providers for a few of its different analysis initiatives.

“Cloud providers are very costly,” Ramsundar mentioned. ​“We’ve discovered that coaching one thing on the size of a big basis mannequin can simply price lots of of 1000’s of {dollars} on the general public cloud. Entry to DOE supercomputing assets makes any such analysis dramatically extra accessible to researchers in trade and academia. Not all of us have entry to the large Google-scale supercomputers.”

Outfitted with 1000’s of graphics processing models (GPUs) and big reminiscence capacities, ALCF’s supercomputers are constructed to deal with the advanced calls for of AI-driven analysis.

“There’s a giant distinction between coaching a mannequin on hundreds of thousands of molecules versus billions. It’s actually not doable on the smaller clusters which are sometimes obtainable to school analysis teams,” Viswanathan mentioned. ​“You simply don’t have the variety of GPUs or the reminiscence wanted to scale fashions to this measurement. That’s why you actually need assets just like the ALCF, with supercomputers and software program stacks designed to assist large-scale AI workloads.”

Nevertheless it’s not simply the computing assets which have propelled this work. The human factor has additionally been important. For the previous two years, Viswanathan’s group has attended the ALCF’s annual INCITE hackathon to work with Argonne computing specialists to scale and optimize their workloads to run effectively on the lab’s supercomputers. 

The challenge has additionally benefitted from collaborations with scientists working to make use of AI in different analysis fields. Argonne computational scientist Arvind Ramanathan, for instance, has been main pioneering analysis in utilizing LLMs for genomics and protein design. Ramanathan, who joined Viswanathan’s INCITE group, has been instrumental in making use of data gained from creating AI fashions for biology functions to battery analysis.

“Everyone seems to be studying from everybody else,” Viswanathan mentioned. ​“Despite the fact that we’re targeted on completely different issues, the innovation stack is analogous. There are these pockets of science, like genomics and chemistry, the place the information has a pure textual illustration, which makes it match for language fashions.”

Reworking the Way forward for Battery Analysis

To make its basis mannequin extra interactive and accessible, the group has built-in it with LLM-powered chatbots like ChatGPT, a novel strategy that’s opening the door to new potentialities for consumer engagement. College students, postdocs and collaborators can ask questions, take a look at concepts rapidly and discover new chemical formulations with no need to put in writing code or run advanced simulations.

“It’s like each graduate scholar will get to talk with a prime electrolyte scientist each day,” Viswanathan mentioned. ​“You’ve got that functionality proper at your fingertips and it unlocks a complete new degree of exploration.”

This functionality can be shifting how researchers take into consideration the invention course of.

“It’s basically altering the best way we’re fascinated by these items,” Viswanathan mentioned. ​“These fashions can creatively assume and give you new molecules which may even make knowledgeable scientists go, ​‘Oh wow, that’s fascinating.’ It’s a rare time for AI-driven supplies analysis.”