‘Discovery learning’ AI tool predicts battery cycle life with just a few days’ data

February 4, 2026
Contact:

A ‘learner,’ ‘interpreter’ and ‘oracle’ work together with minimal experiments to draw parallels between historical data and new battery designs

At the start of the flow chart, a gray 3x3x3 cube built from smaller constituent cubes is labeled "input space". The arrow points to the "Learner (active learning)" then flows out to query selection, represented as a gray 3x3x3 cube with five constituent cubes colored yellow and pulled halfway out from the gray assembly. A yellow cube is labeled "selected samples" and points to the "Interpreter (physics-guided learning)." "Distribution-shifted yet zero-cost historical datasets," represented as blue and green cylinders, also flow to the Interpreter via a dotted line. The next arrow is labelled "Space alignment" pointing to a transparent cube with three constituent cubes shown in yellow, blue, and green, labelled "Universal feature space." The next arrow is labelled "training" and goes to the "Oracle (zero-shot learning)." A pink constituent cube emerges from the Oracle, labeled "primary inference," which flows into a new version gray cube with the constituent cubes that had previously been yellow now shown in pink. An arrow labeled "training" leads back to the Learner (active learning). A purple constituent cube emerges from the Learner labelled "secondary inference," and the final 3x3x3 cube is composed of pink and purple constituents.
A diagram of how the discovery learning system developed at U-M works. The “learner” is presented with a new battery design, for which it must figure out the cycle life. The “learner” identifies some battery designs that would fill in gaps in its knowledge, needed to make the prediction. These batteries are then built and tested for tens of cycles, and the data goes to the “interpreter.” The interpreter analyzes the experimental results with physical models. It extracts key, physically understandable attributes that generalize across both historical datasets and new battery designs. The “oracle” then leverages the most relevant information to make cycle life predictions for the prototypes tested for tens of cycles. The learner then combines that information with what it knows from earlier predictions, producing the lifetime prediction for the unbuilt prototypes. Image credit: Jiawei Zhang, Ziyou Song Group, University of Michigan. Copyright: Nature Publishing Group

An agentic AI tool for battery researchers harnesses data from previous battery designs to predict the cycle life of new battery concepts. With information from just 50 cycles, the tool—developed at University of Michigan Engineering—can predict how many charge-discharge cycles the battery can undergo before its capacity drops below 90 percent of its design capacity.

This could save months to years of testing, depending on the conditions of cycling experiments, as well as substantial electrical power during battery prototyping and testing. The team estimates that the cycle lives of new battery designs could be predicted with just 5% of the energy and 2% of the time required by conventional testing.

Ziyou Song
Ziyou Song

“When we learn from the historical battery designs, we leverage physics-based features to construct a generalizable mapping between early-stage tests and cycle life,” said Ziyou Song, U-M assistant professor of electrical and computer engineering and corresponding author of the study in Nature. “We can minimize experimental efforts and achieve accurate prediction performance for new battery designs.”

The study was funded by the battery company Farasis Energy USA in California, which also provided battery cells and data from its design and testing to assess how well the model—trained only on free, public data—performed.

The tool is inspired by a teaching approach known as discovery learning, or learning by doing. A student learning in this way has a problem to solve and resources to help discover the solution, while drawing on their own experiences and prior knowledge. Over the course of solving many problems, the student no longer needs the resources to solve similar ones—they have internalized the knowledge and skills.

Jiawei Zhang
Jiawei Zhang

“Discovery learning is a general machine-learning approach that may be extended to other scientific and engineering domains,” said Jiawei Zhang, U-M doctoral candidate in electrical and computer engineering and the first author of this study, who had the initial inspiration to design a team of AI agents that could simulate this mode of learning.

How the AI discovery learning tool works

The team designated an AI “learner” that would predict the cycle life for a given battery design and cycling conditions, such as temperature and current. The learner chooses a few battery candidates that would fill gaps in its knowledge, to be built and run for about 50 cycles. The results of those experiments flow to an “interpreter,” which accesses historical data and runs calculations with a physics-based battery simulator. The “oracle” then makes cycle life predictions for the experimental batteries based on the historical data and calculations provided.

In the conventional process, six new designs are manufactured, yielding six prototypes, which then enter life testing with a long dotted line indicating time, finally leading to a graph showing the match with performance criteria. In the discovery process, two prototypes are manufactured, followed by a short testing phase that yields an initial prediction of performance criteria, enabling predictions for four remaining unbuilt prototypes, as a result. The arrow from the performance criteria feeds back to the design and prototyping.
A diagram of how prototyping and testing work in conventional battery design, with months to years of lifetime testing, and with discovery learning, which enables designers to iterate after just days to weeks of testing. The U-M researchers hope that their system helps speed the development of longer-lasting batteries, and that their “discovery learning” framework finds uses in other areas of engineering as well. Image credit: Jiawei Zhang, Ziyou Song Group, University of Michigan

Finally, the learner combines the new information with previous predictions to estimate the cycle life of the new battery design. Even with experiments, the discovery learning system provides huge time and energy savings, with the potential to improve further as the learner accumulates enough knowledge to make predictions without running the discovery loop.

Next-gen lithium-ion batteries are very different from previous iterations—in chemistry, structure and materials—but the team argues that there are parallels among them that may help predict how new designs will perform. Rather than using simple statistical features from current and voltage signals, the interpreter leverages underlying physical properties to establish commonalities among different batteries.

The illustration depicts how the AI tool built at U-M draws on tests of earlier battery cells to predict the cycle life performance of new battery designs. The glowing cells on each cube represent the way the system identifies gaps in its knowledge and fills them with short experiments, physics-based computer modeling, and historical data. Image credit: Xin Zou, University of Wisconsin–Madison.
The illustration depicts how the AI tool built at U-M draws on tests of earlier battery cells to predict the cycle life performance of new battery designs. The glowing cells on each cube represent the way the system identifies gaps in its knowledge and fills them with conclusions drawn from historical data, physical models, and data-driven models. While the historical data set is exclusively cylindrical cells, the tool correctly predicted the lifetimes of pouch cells as well. Image credit: Xin Zou, University of Wisconsin–Madison.

With this information in hand, the oracle considers the battery in two ways: its internal characteristics—information from the interpreter about the physics and chemistry of the cell—and its operating conditions. For instance, at higher temperatures, a particular chemical change may dominate how the battery is likely to degrade, but that mechanism is less important at lower temperatures.

The team tested out their model with data and pouch cells from Farasis Energy USA. After training on a data set that included only cylindrical cells, similar to the familiar AA battery, the model could predict the performance of these larger cells. While full tests run to 1,000 cycles and can take a few months to years, 50-cycle tests take only a few days to weeks, according to the team’s estimates. Testing required fewer cells, as well as fewer cycles, resulting in energy savings of about 95%.

Within battery technology, the team intends to expand the approach to other areas of performance, such as safety and charging speed. However, as discovery learning is a new scientific machine-learning approach, the team believes that others could build similar predictive tools or develop new approaches to optimization. They hope it could speed development in many disciplines bottlenecked by the need for expensive experiments, most immediately in chemistry and material design.

Researchers from the National University of Singapore also contributed to the study.