Combining Physics and Machine Learning to Predict Battery Aging with Confidence

Battery technology is evolving rapidly to meet the growing demands of electric vehicles, large-scale energy storage systems, and portable electronics. A major challenge lies in reliably predicting long-term battery performance within practical development timelines. Because batteries degrade gradually during both use and storage, conventional testing methods take a long time to produce accurate lifetime estimates. These lengthy evaluation cycles slow down the process of validating new designs, assessing durability, and defining warranty limits. For battery engineers, this delay translates directly into longer development cycles, higher validation costs, and conservative design margins. This creates a need for faster and more predictive approaches to battery aging assessment.

Accurate prediction of battery aging is essential for effective design, testing, and system integration. Traditional experimental approaches are often time-consuming, while high-fidelity simulations, though precise, require significant computational resources. By combining physics-based tools like GT-AutoLion with machine learning (ML), virtual aging datasets can be generated to train machine learning metamodels. A metamodel represents a complex physical system model as a reduced mathematical form. These models enable engineers to predict battery state of health (SOH) much more rapidly. This blog explains how this workflow accelerates both calendar aging and cycle aging predictions, reducing or even eliminating the need for long-duration experiments.

Why Machine Learning for Battery Aging?

Classical physics-based battery models, such as the Pseudo-2D model, provide high reliability and precision but are often computationally expensive, especially when many scenarios need to be evaluated. Engineers are constantly looking for ways to accelerate these processes without sacrificing accuracy. Machine learning offers a promising solution: by creating a surrogate model, outputs can be predicted rapidly after training. ML models are particularly well-suited for aging applications, where results are sampled at intervals, such as after a few storage days or each cycle.

In many cases, engineers may not have extensive measurement data, especially during early development stages. This is where GT-AutoLion, a physics-based model that includes detailed aging mechanisms, comes into play. Once the model is calibrated against a limited set of measurements, it can generate large-scale virtual testing data across a wide range of operating conditions. This dataset becomes the foundation for training a reliable ML metamodel.

With the help of the ML Assistant tool in GT-SUITE, engineers can easily import data, whether from GT-Post, csv, Excel, TXT, or MAT files, and train metamodels for fast and accurate aging prediction. These models are useful for studying various types of battery aging, such as calendar and cycle aging.

To illustrate how this physics-based machine learning workflow works in practice, let’s look at two representative aging scenarios:

Case 1: Predicting Calendar Aging Using the ML Assistant Tool Within GT-SUITE

Calendar aging refers to the loss of battery capacity that occurs while the battery is resting, without any charge or discharge cycles. Even when not in active use, chemical reactions inside the cell continue, slowly reducing its capacity. The rate of this degradation is strongly influenced by factors such as state of charge (SOC) and temperature, making these key parameters for the calendar aging predictive model.

Generating virtual calendar aging data

To build a ML metamodel, we first used GT-AutoLion to generate virtual storage data based on a Design of Experiments (DoE). The DoE varied parameters such as state of charge (SOC) and initial temperature. GT-AutoLion then simulated long-term storage for up to five years, recording SOH values at regular intervals (e.g., every 30 days). SOC and temperature are employed as the input factors, while SOH values are used as outputs for training the ML model.

Training and testing the metamodel

The key challenge in predicting calendar aging is extrapolation, estimating degradation beyond the period covered by the training data. To determine the minimum fraction of storage days needed for effective training, we studied different subsets of the storage days, including 20%, 30%,…, and 60%. The best results were obtained using 60% of the total storage days for training, with the remaining 40% reserved to evaluate the model’s capability to predict future aging. The results, illustrated in Figure 1, highlight the metamodel’s ability to:

• achieve strong accuracy within the trained region
• demonstrate excellent extrapolation capability beyond the training range
• maintain close agreement with the physics-based GT-AutoLion reference results

Figure 1. A comparison of calendar aging prediction results using the ML Assistant tool versus the physics-based model using Gamma Technologies’ GT-AutoLion

Figure 1 shows that the machine learning model closely matches the physics-based results and successfully preserves the nonlinear degradation behavior captured by the physics-based model.

In practice, this means engineers can confidently estimate five years of degradation even if the metamodel is trained on just the first two to three years of data. This dramatically reduces development time, enabling faster iteration and earlier design decisions.

Case 2: Predicting cycle aging using the ML assistant tool within GT-SUITE

While calendar aging measures degradation at rest, cycle aging captures degradation from repeated charge–discharge cycles. This form of aging depends on several operating variables, including temperature, depth of discharge, and charge/discharge C-rates.

Virtual cycle aging simulations

Using GT-AutoLion, we simulated battery aging across a wide range of conditions. Each simulation is followed by a consistent cycling protocol designed to estimate capacity. The protocol began with a 10-minute rest period to allow the cell to reach equilibrium, followed by a constant-current (CC) discharge to deplete stored energy and a constant-current, constant-voltage (CCCV) charge to restore capacity. This cycle was repeated under varying conditions until the battery reached end-of-life (defined as 80% SOH). Inputs for the ML model include initial temperature, cycle number, charge C-rate, discharge C-rate, and depth of discharge, while the output is SOH, a scalar variable used for ML modeling.

Training and extrapolation results

To evaluate the metamodel extrapolation capability for cycle aging, we wanted to determine the minimum number of cycles required for effective training. We examined different numbers of cycles and found that using the first 300 cycles of each simulation for training provided the best results. The remaining cycles were reserved for testing the model to predict future degradation.

Figure 2. A comparison of cycle aging prediction results using the ML Assistant tool versus the physics-based model (GT-AutoLion)

Just like in the calendar aging case, the ML model performed significantly well, as shown in Figure 2:

• It reproduced the GT-AutoLion degradation trends with high accuracy
• It reliably predicted late-stage aging data not included in the training set
• It generalized across different operating conditions

This means accurate lifetime cycle-aging predictions can be made using only a fraction of the total cycles, significantly reducing simulation and validation effort.

Conclusion: Accelerating Battery Design via Machine Learning-physics Integration

Both case studies show an exciting trend in battery engineering: combining physics-based modeling with ML. GT-AutoLion creates detailed virtual data, and ML models turn that data into fast, easy-to-use tools that can predict beyond what they were trained on. This approach dramatically shortens development cycles, reduces reliance on lengthy experiments, and delivers accurate insights with minimal computational resources.

Learn More About ML-Accelerated Simulation

Discover how integrating physics-based models with ML metamodels can significantly reduce simulation time while preserving fidelity. Read our blog on Dynamic Machine Learning for Modeling and Simulation to explore real-world applications, and visit our Machine Learning and Optimization Solutions page for more related content. Join our LinkedIn community to stay updated on simulation-driven workflows, battery modeling insights, and upcoming technical content. You can also contact us to learn how Gamma Technologies can support your machine learning goals.