Understanding the complexity of battery health with explainable AI
By: Shijing Sun, Amalie Trewartha, and the TRI Energy & Materials research division
Just as we all care about our health and quality of life over time, battery scientists have the same passion for maximizing the longevity of batteries and closely tracking their health. The health of a battery, just as our own, is a complex product of internal and external factors that cannot be easily summarized by a single number.
A battery is a composite device, where each of the subcomponents, such as the anode and cathode, ages in different ways. The wide variety of conditions in which a battery is used also comes into play: the batteries in a hybrid vehicle driven in Arizona face very different demands than a fully electric vehicle in Alaska. To see a complete picture of battery health, scientists are required to understand the diverse ways in which batteries can age and methods to link a battery’s degradation to its design and usage.
Battery research challenges
The complexity of battery aging makes it difficult to study. As scientists, we can use our knowledge of physics and chemistry to develop an understanding of different aging mechanisms, but the wide range of scenarios and conditions for battery usage makes it difficult for us to gain a holistic perspective quickly. This makes it essential to draw out the subtle correlations from large battery datasets, which can then inform the judgment of scientists. Just as we use large datasets and advanced analytics to understand what helps people to maintain good health with age, we need to do the same to understand battery health.
The use of artificial intelligence (AI) on data from batteries, a field known as battery informatics, has rapidly emerged as an important part of the battery scientist’s toolkit. Previous work from TRI in collaboration with Stanford and MIT has shown that AI has the ability to accurately predict the lifetime of batteries and develop optimized algorithms for fast charging. While the tools of conventional AI are undoubtedly powerful for making predictions, they’re often “black box” models, with limited ability to offer explanations for how and why those predictions are made. This can make it difficult for researchers to draw wider conclusions or gain scientific insight, so in order to continue building on the success of our previous results, it requires us to adopt new approaches – such as adding explainability to predictive power.
AI-driven solutions
Two novel approaches were introduced to enable interpretability in our recent work here. First, our team used an interpretable machine learning technique, SHAP analysis, which originates from game theory. SHAP analysis allows us to quantify the contributions from individual inputs to a machine learning model. For example, it allows us to deduce whether increasing the charging speed or using a wider range of the battery’s capacity is more detrimental to battery health. Second, inputs to the AI models are restricted to features developed by our team’s battery experts that are physically meaningful and interpretable. Sixteen metrics were developed to describe the state of health of batteries, ranging from the cell-level to electrode-specific descriptors. As a result, the output of our predictive models can be clearly mapped to scientific understanding.
While developing our explainable AI approach, researchers at TRI worked with counterparts at Stanford and MIT over a period of two years to collect a dataset of aging trajectories of 363 commercial EV battery cells. This collaboration resulted in a comprehensive and diverse dataset, containing a wealth of information about the state of health of the batteries tested under 218 distinct aging conditions. The range of aging conditions leads to battery lifetimes ranging from 4600 cycles for the gentlest conditions to 63 cycles for the harshest. To put that into perspective, for an EV that’s charged daily, that’s the difference between a battery that lasts 12 years to one that lasts just two months!
Being able to predict and explain the complex state of health of batteries provide tremendous value to battery engineers looking to improve the current battery design, as well as for policymakers to develop roadmaps of charging and usage guidance for EV drivers. The success of this work exemplifies a new way of conducting scientific research, where AI acts as a co-worker to scientists. Instead of viewing AI as a replacement for humans, here at TRI, we develop AI toolkits that work hand-in-hand with humans to enhance the capabilities of both.
At TRI, we are focused on research with the potential to improve the quality of life for individuals and society. And we are trying to solve one of the biggest challenges of our time: how do we power our future in a sustainable, affordable way? TRI’s Energy and Materials team is dedicated to researching and developing new methods to discover new materials and improving batteries to help humanity achieve these goals. For more on TRI’s approach to AI for science, learn more here (outlined in a previous Medium article).
