See all NewsEngineering News

AI Algorithm Identifies High-Performing Electrolytes for Batteries

New tool provides a solution to an intractable problem in the chemical space

The Problem

Developing longer-lasting, more-efficient electrolytes better than those engineered more than three decades ago has been difficult.

Our Idea

An AI model that uses machine learning to rapidly predict the properties of electrolyte mixtures.

Why It Matters

The AI algorithm can significantly speed up and guide the discovery of more efficient and longer-lasting electrolytes for batteries.

Our Team

Professor Wei Chen, Professor James Rondinelli, PhD student Henry Zhang

Electrolytes enable the batteries in everything from your cell phone to an electric car to generate power. Scientists have long sought to develop longer lasting, more efficient electrolytes, but still feel challenged to discover any that perform better than those engineered more than three decades ago. One major reason for this: the vast number of potential molecular combinations that can create an electrolyte. 

Wei Chen, James Rondinelli

Scientists at Northwestern Engineering in partnership with researchers at the University of Texas at Austin have developed an artificial intelligence (AI) algorithm called MolSets that could help researchers move past this issue. The AI model uses machine learning to predict the properties of electrolyte mixtures, using the power of computation to complete in seconds a process that would take years to do in the lab.

“Imagine that you want to make a new molecular mixture and you need to select three molecules out of 200, that’s more than a million potential candidates – not to mention molecules can have varying different weight fractions. You can’t test all of those in the lab,” said Henry Zhang, a PhD student and Ryan Fellowship recipient at Northwestern Engineering and first author of the paper. “We created a model that can predict the properties of a molecular mixture before you make it. This can help make the search for new electrolytes more guided and efficient.”

A research paper about the findings, titled “Learning Molecular Mixture Property Using Chemistry-Aware Graph Neural Network,” was published June 12 in PRX Energy.

Combating combinatorial complexity

To develop a workable AI model, the researchers first needed to tackle a challenge in the chemical space known as combinatorial complexity. In molecular mixtures this occurs not only in the sheer number of potential combinations, but also in the multitude of ways molecules interact with each other. The algorithm also needed to have permutation invariance, meaning that the results stay the same even when the order of the molecules input changes.

The solution was to represent individual molecules as graphs and their mixture as a mathematical set and then leverage a graph neural network and “deep sets” architecture to extract information at the molecular level and aggregate it at the mixture level. This addressed local complexity while retaining global flexibility.

“Think of it like ChatGPT, but for molecular mixtures,” Zhang said. “Just like ChatGPT captures the relation between words, our algorithm captures the interactions between molecules.”

The team trained the model using a molecular mixture dataset collected from published scientific papers, and the University of Texas at Austin researchers experimentally validated the algorithm’s mixture predictions.

Think of it like ChatGPT, but for molecular mixtures. Just like ChatGPT captures the relation between words, our algorithm captures the interactions between molecules.

Henry ZhangPhD Student

Having achieved accuracy, efficiency and interpretability with the algorithm, the team now plans to further refine it with the addition of more data. Just like ChatGPT benefits from the number of words and phrases inputted, MolSets benefits from the number of molecular datasets. In the future, they see their model as part of a “self-driving laboratory” that uses robotics to autonomously generate massive amounts of experimental data and use artificial intelligence (AI) to guide automated experiments.

“AI-guided experimentation is going to be transformative and become a foundational component to future materials and chemistry research,” said James Rondinelli, co-corresponding author of the paper and Walter Dill Scott Professor of Materials Science and Engineering at Northwestern Engineering. “It’s both cost saving and time saving. And once a model has been trained, it is available to everyone, which in a sense brings equity to the research and engineering enterprise and hopefully accelerates it societal impact.”

This work also serves as an example of the interdisciplinary research that defines Northwestern Engineering’s approach to scientific discovery.

“There are challenges within disciplines that cannot be addressed by the conventional methods of those disciplines. This work, for example, is a collaboration between materials and mechanical engineering researchers that helped advance the field of materials design using artificial intelligence,” said Wei Chen, co-corresponding author of the paper and Wilson-Cook Professor in Engineering Design and chair of mechanical engineering at Northwestern Engineering. “We wouldn’t have had that impact without developing collaborations and partnerships across disciplines. Without interdisciplinary collaboration or an interest in interdisciplinary thinking, we wouldn't be able to identify certain challenges or know that those challenge are addressable.”

Additional authors on the paper include Jie Chen of Northwestern Engineering; and Tianxing Lai and Arumugam Manthiram of the University of Texas at Austin.