When Microsoft researchers in 2023 identified a new kind of material that could dramatically reduce the amount of lithium needed in rechargeable batteries, it felt like combing through a haystack in record time. That’s because their discovery began as 32 million possibilities and, with the help of artificial intelligence, produced a promising candidate within 80 hours.
Now researchers at the Pacific Northwest National Laboratory plan to synthesize and test the novel material, NaxLi3−xYCl6, in a battery setup. It’s one of several AI-generated battery chemistries making its way to the real world.
Microsoft’s experiment started when the researchers wanted to demonstrate how AI could tackle the needle-in-a-haystack problem of finding useful new materials and chemicals. They decided to seek new candidates for a rechargeable battery’s electrolyte, because a better electrolyte could make batteries safer while simultaneously improving performance, says Nathan Baker, project leader at Microsoft for Azure Quantum Elements, a program to accelerate chemistry and materials research through Microsoft’s advanced computing and AI platforms.
“Our goal was to take one of these AI models and show the promise of accelerating scientific discovery—sifting through 32.5 million materials candidates and showing that we could do it in a matter of hours, not years,” Baker says. Their model, called the M3GNet framework, accelerated simulations of molecular dynamics to evaluate properties of the materials such as atomic diffusivity.
First, the Microsoft researchers asked the model to drop new chemical elements into known crystalline structures in nature and determine which resulting molecules would be stable, a step that cut the 32 million starting candidates down to half a million. AI then screened those materials based on the necessary chemical abilities to make a battery work, which chopped the pool to just 800. From there, traditional computing and old-fashioned human expertise identified the novel material that could function within a battery and use 70 percent less lithium than the rechargeable batteries in commercial use today.
AI’s Role in Next-Gen Battery Design
The Microsoft team isn’t alone. Around the world, researchers are busy trying to develop next-generation designs to replace or improve lithium-ion batteries, which use large quantities of rare, expensive, and difficult-to-acquire elements. New battery designs could use more abundant materials, reduce the fire danger from lithium-based liquid electrolytes, and pack more energy into a smaller space. The chemistries to do this are waiting out there to be discovered, and increasingly, researchers are harnessing AI and machine learning to do the work of sorting through the mountain of data.
“We are teaching AI how to be a materials scientist,” says Dibakar Datta, associate professor at the New Jersey Institute of Technology, who published a study in August that used AI to identify five candidate materials for batteries that would outperform Li-ion. Datta’s team is working on the multivalent battery: one that employs multivalent ions that can carry multiple charge levels as opposed to the single charge carried by a lithium battery.
This would give the battery a greater energy storage capacity, but it also means working with larger ions from elements higher on the periodic table, like magnesium and calcium. Those larger ions won’t necessarily fit into existing battery designs without cracking or breaking the elements, Datta says. His new study used what he calls a crystal diffusion variational autoencoder (CDVAE) that could propose new materials, and a large language model (LLM) that could find materials that would be the most stable in the real world. From a pool of millions of possibilities, the approach found five porous materials of the right size that could do the job.
Guiding an AI model on its hunt through the nearly infinite space of possible materials is the tipping point in this field. The key to using it as a research partner is to find a happy medium between a model that works fast and a model that delivers perfectly accurate results, says Austin Sendek, professor at Stanford University who has developed algorithms to help AI discover new battery materials.
“You have to traverse both breadth and depth,” says Sendek. Depth, because designing these things takes a lot of deep scientific knowledge about properties, engineering and chemistry, and breadth, because you have to apply that knowledge across an infinite chemical space, he says. “That’s where the promise of AI comes in.”
AI Battery Technology Search at IBM
Researchers at IBM have taken an AI-driven approach to identify new electrolyte candidates, which involved identifying chemical formulations with far higher ionic conductivity than the lithium salts used in current batteries. A typical electrolyte can contain six to eight ingredients including salts, solvents, and additives, and it’s nearly impossible to consider all the combinations without AI.
To whittle down the field, the IBM team developed chemical foundation models trained on billions of molecules. “They capture the basic language of chemistry,” says Young-Hye Na, Principal Research Staff Member at IBM Research. Her team then trains those models with battery-related data so the AI can predict important properties for battery applications on scales from individual molecules all the way up to a whole device. Na described the work in a paper published in August in NPJ Computational Materials.
Because the work investigates new combinations of existing materials rather than using AI to invent exotic new materials, its potential to help build the battery of tomorrow is that much more promising, Na says. The IBM team is now collaborating with an undisclosed EV manufacturer to design high-performance electrolytes for high-voltage batteries.
IBM’s use of AI for batteries isn’t limited to the hunt for promising materials. Typically, when AI reveals a promising new material, the next step is for experimentalists to synthesize the stuff, experiment with it in the lab, and one day to test it in a real device. Machine learning (ML) will aid researchers in this testing step, too.
IBM is testing the real-world viability of new battery setups by building their digital twins—virtual models that allow the researchers to predict how a particular battery chemistry would degrade over a lifetime of countless power cycles. The model, developed in collaboration with battery startup Sphere Energy, can predict a battery’s long-term behavior in as few as 50 power cycles modeled on the digital twin, says Teodoro Laino, distinguished research staff member at IBM Research.
The next phase of AI battery research is quantum. As Microsoft and IBM push toward the potential of quantum computers, both see its promise to model complex chemistry with no shortcuts or compromises. Na says that while current AI is a crucial tool for investigating battery chemistry, the next step—modeling whole EV battery packs, for example, and taking into consideration all the variables they encounter in the real world—would require the power of quantum computing.
As Baker puts it: “We know classical computers have problems generating accurate answers for complex substances, complex molecules, complex materials. So our goal right now is actually to change the way the data is generated by bringing quantum into the loop so that we have higher accuracy data for training ML models.”
This article was updated on October 2, 2025.
From Your Site Articles
Related Articles Around the Web