From Months to Days:
AI-Predicted Battery Cell End-of-Life
Leading Global Automotive OEM
The situation
Validating battery cells through physical testing is extremely time- and resource-intensive. A standard end-of-life (EoL) validation process for a single cell can take months, or even years. The client needed to make better use of historical data and reduce the number of physical Design-of-Experiments (DOE) test series required.
The challenge
How vetta changes the process
vetta combines historical tests, simulations and new validation data into a single battery intelligence model. Weeks of testing become years of insight.
The approach
A Foundation Model for Battery Aging — Tuned to the Client's Cells
Sphere used VETTA, its time-series foundation model pre-trained on 10+ million synthetic datasets and thousands of real cell tests. The approach comprised:
Model Fine-Tuning
The pre-trained model was refined using the client's specific cell data.
Hybrid Simulation
A combination of data-driven AI and physics-based modeling (P2D models) to not only predict but physically explain varying degradation mechanisms, such as SEI growth, lithium plating, etc.
Defined Use Cases
Three specific use cases were defined: DOE optimization for 2 cell types, and transfer learning to a new cell for further validation.
What changes for the engineer
Instead of waiting months for a full DOE test campaign, the engineer feeds the first few weeks of cycling data into VETTA and receives an end-of-life prediction — together with the physical degradation drivers behind it — within days. This supports earlier go/no-go decisions and a smaller, more targeted physical test matrix.
The results
The model moved directly from proof of concept to a validated prediction tool, integrated into the client's cell validation workflow across multiple cell types and projects.
Prediction accuracy of ~ < 3% MAPE achieved for DOE optimization
Zero-shot transfer to new cell types achieved under 3% error, improving to under 2% MAE with just 150 cycles of data
Degradation mechanisms confirmed — predictions aligned with known physical drivers such as SEI growth and lithium plating
Early test termination enabled by flagging tests likely to miss requirements before they run to completion
