To help combat climate change, many car manufacturers are racing to add more electric vehicles to their lineups. But to convince prospective buyers, manufacturers need to improve how far these cars can go on a single charge. One of their main challenges? Figuring out how to make extremely powerful but lightweight batteries.
Typically, however, it takes decades for scientists to thoroughly test new battery materials, says Pablo Leon, an MIT graduate student in materials science. To accelerate this process, Leon is developing a machine-learning tool for scientists to automate one of the most time-consuming, yet key, steps in evaluating battery materials.
With his tool in hand, Leon plans to help search for new materials to enable the development of powerful and lightweight batteries. Such batteries would not only improve the range of EVs, but they could also unlock potential in other high-power systems, such as solar energy systems that continuously deliver power, even at night.
Leveraging machine learning to research battery materials
Scientists investigating new battery materials generally use computer simulations to understand how different combinations of materials perform. These simulations act as virtual microscopes for batteries, zooming in to see how materials interact at an atomic level. With these details, scientists can understand why certain combinations do better, guiding their search for high-performing materials.
But building accurate computer simulations is extremely time-intensive, taking years and sometimes even decades. “You need to know how every atom interacts with every other atom in your system,” Leon says. To create a computer model of these interactions, scientists first make a rough guess at a model using complex quantum mechanics calculations. They then compare the model with results from real-life experiments, manually tweaking different parts of the model, including the distances between atoms and the strength of chemical bonds, until the simulation matches real life.
With well-studied battery materials, the simulation process is somewhat easier. Leon says that scientists can buy simulation software that includes pre-made models, but these models often have errors and require additional tweaking.
To build accurate computer models more quickly, Leon is developing a machine-learning-based tool that can efficiently guide the trial-and-error process. “The hope with our machine learning framework is not to have to rely on proprietary models or do any hand-tuning,” he says. Leon has verified that his tool is as accurate as the manual method for building models for well-studied materials.
With this system, scientists will have a single, standardized approach for building accurate models instead of the patchwork of current approaches, Leon says.
Leon’s tool comes at an opportune time when many scientists are investigating a new paradigm of batteries: solid-state batteries. Compared to traditional batteries containing liquid electrolytes, solid-state batteries are safer, lighter, and easier to manufacture. But creating versions of these batteries that are powerful enough for EVs or renewable energy storage is challenging.
This is largely because in battery chemistry, ions dislike flowing through solids and instead prefer liquids, in which atoms are spaced further apart. Still, scientists believe that solid-state batteries can provide enough electricity for high-power systems, such as EVs, with the right combination of materials.
Leon plans to use his machine-learning tool to help look for good solid-state battery materials more quickly. After finding some powerful candidates in simulations, he’ll work with other scientists to test the new materials in real-world experiments.
Source: MIT News Office