Researchers have developed a multi-fidelity framework for lithium-ion battery lifespan prediction that combines coupled degradation mechanisms with machine learning, aiming to improve early prediction of both battery knee point and remaining useful life. The work addresses a major challenge in battery management: how to make reliable forecasts before long-term aging data have accumulated, when operators still have time to act on the results.
Accurate lifespan prediction matters because battery systems are often used in applications where safety, reliability, and economic efficiency all depend on knowing how degradation will evolve. If a battery's remaining useful life, or RUL, can be estimated too late, it becomes harder to optimize operation, plan maintenance, prevent failures, or decide when a pack is suitable for continued service or second-life use. Yet early prediction is notoriously difficult. Battery aging is driven by multiple interacting processes, and the most useful warning signals may emerge long before enough empirical degradation data exist to train a conventional purely data-driven model with confidence.
The authors of the new study point out that existing methods each have important limitations. Model-based methods can incorporate electrochemical knowledge, but they often become computationally expensive and cumbersome in practical deployment. Data-driven methods can fit observed degradation patterns, but they may struggle when aging data are sparse and may not offer enough mechanistic interpretability to support confident engineering decisions. Real-world battery prognosis often suffers from exactly this combination of constraints: limited early-cycle data, strong nonlinear degradation, and the need for predictions that are both accurate and physically meaningful.
To address that gap, the researchers proposed a multi-fidelity framework that explicitly combines physics-informed degradation understanding with sparse early-life measurements. The study first develops an enhanced pseudo two-dimensional, or P2D, battery model that incorporates both solid electrolyte interphase growth and partially reversible lithium plating dynamics. This matters because many conventional models do not fully capture the coupled nonlinear degradation patterns that shape how capacity fades over time. By embedding these degradation pathways in the mechanistic model, the framework is able to represent battery aging in a way that more closely reflects the physical causes of performance loss.
The work then goes a step further by examining how specific mechanistic parameters influence the shape of the capacity decay trajectory. According to the paper, three governing parameters were identified as especially important: SEI solvent diffusivity, the lithium-plating kinetic rate constant, and the dead-lithium decay constant. By linking those parameters to observed aging behavior, the study establishes quantitative connections between molecular-scale degradation processes and macroscopic battery performance decline. That is a significant contribution because it helps explain not only that a battery is aging, but why the degradation curve is taking a particular form.
This mechanistic understanding is then fused with data in a multi-fidelity regression architecture designed to predict both knee point and RUL. The battery knee point is especially important because it marks the transition into a phase of more rapid capacity decline, where battery usefulness can deteriorate much faster. Predicting that transition in advance can be highly valuable for safety management and lifecycle planning. The new framework is designed to do this even when only sparse early-life capacity measurements are available, using physics-based degradation trajectories to provide structure that the limited data alone may not contain.
According to the article, the framework was validated across 169 commercial cells subjected to different fast-charging protocols. The reported results show Mean Absolute Percentage Errors of 5.7% for predictions of knee point and RUL. Those numbers are meaningful because they suggest the method can maintain useful accuracy even under varying fast-charging conditions, which are particularly relevant to modern battery applications. Just as importantly, the paper emphasizes that the framework achieves this level of prediction while reducing the need for extensive late-life aging data, making the method more practical for early prognosis.
Taken together, the study suggests that battery prognosis may benefit most when physics and machine learning are not treated as competing approaches, but as complementary sources of information. More validation will still be needed across broader chemistries, use conditions, and field datasets. Even so, the framework offers a strong example of how mechanistic degradation modeling can be fused with sparse data to improve interpretability and prediction at the same time. For battery systems used in transport, energy storage, and other demanding applications, that could support safer operation, longer service life, and better-informed decision-making long before a battery reaches its end of life.
Reference
Author:
Dong Lu a , Fei Ren a , Changlong Li a , Haoyu Ming a , Naxin Cui a
Title of original paper:
Fusing Coupled Degradation Mechanisms with Machine Learning: A Multi-fidelity Framework for Lithium-ion Battery Lifespan Prediction
Article link:
https://www.sciencedirect.com/science/article/pii/S2773153725001173
Journal:
Green Energy and Intelligent Transportation
DOI:
10.1016/j.geits.2025.100367
Affiliations:
School of Control Science and Engineering, Shandong University, Jingshi Road 17923, Jinan 250061, PR China
Experimental study
Not applicable
4-Mar-2026