The last several years have produced a significant amount of published work on applying machine learning to cycle life prediction for lithium-ion batteries. The approach is attractive: gather charge-discharge cycle data from cells aging under various conditions, extract features from the voltage and current traces, train a model to predict end-of-life cycle count from early-cycle data, deploy to infer remaining useful life in the field. Done well, this approach can produce predictions that outperform simple empirical models calibrated on historical data.
Done poorly, which happens more often than the publication record suggests, the models are accurate on the training distribution and fragile outside it. The core issue is feature selection. Raw voltage traces and simple statistics computed from them (discharge capacity, mean voltage, capacity variance) can capture variance in the training set without capturing the underlying physical mechanisms. A model that learns to predict cycle life from a raw voltage trace feature may be learning a correlation that exists in the training data but does not generalize to cells of a different chemistry, age profile, or operating protocol.
Physical degradation mechanisms as the feature design constraint
The alternative to feature-as-statistics is feature-as-physics: design features that directly quantify the progress of known degradation mechanisms. For lithium-ion cells, the primary mechanisms affecting cycle life are SEI growth, lithium plating and stripping, cathode particle cracking under cycling stress, and electrolyte decomposition. Each of these processes has observable signatures in the electrochemical data.
SEI growth is the most widely characterized mechanism in NMC/graphite cells. It consumes lithium irreversibly (showing up as declining Coulombic efficiency), increases the ohmic resistance of the anode-electrolyte interface (showing up in impedance), and produces a small but measurable amount of gas in some electrolyte systems (not directly accessible from electrical measurements in most applications). The Coulombic efficiency at cycle n, particularly the deficit between charge capacity and discharge capacity, is a physically meaningful SEI growth proxy. A model feature that normalizes this deficit against cycle count and temperature history is physically grounded in a way that raw voltage statistics are not.
Lithium plating signatures appear in the differential capacity (dQ/dV) curves as secondary plateaus or shoulder features in the 3.5 to 3.8V region during low-rate characterization cycles. These features come from the stripping of lithium metal that was deposited during a previous high-rate charge event. A model that includes dQ/dV feature position and magnitude as inputs is encoding actual physical information about the extent of prior plating events.
The training data problem
Even with physically grounded features, the training data challenge for cycle life prediction is severe. Cycle life datasets require cells to be cycled to end-of-life, which takes months to years depending on the protocol. The existing public datasets (several of which come from the Severson et al. work at Stanford and MIT, now widely cited) cover a limited range of cell chemistries, form factors, and cycling protocols. Cells cycled at 45 degrees Celsius ambient age differently than cells cycled at 25 degrees Celsius. Cells charged at 4C age differently from cells charged at 1C. A model trained on 25 degree / 1C data does not transfer cleanly to Indian two-wheeler packs cycled at 35 degree ambient with variable charge rates.
We are not claiming that existing public data is useless for this application. It is valuable as a source of prior knowledge about degradation mechanism signatures and the relative importance of different features. The issue is that the specific quantitative predictions from models trained on that data should not be trusted without validation against cells from the actual application chemistry and operating conditions. The features may transfer better than the calibration coefficients.
For a pack engineering team building a cycle life prediction capability for a specific cell and operating environment, the practical path is: gather cell cycling data from early characterization under representative operating conditions (even 50 to 100 cycles is useful for feature calibration), use physically-grounded features that have been validated in the literature as mechanistic proxies, and build a model that is parameterized on your data rather than exclusively on public datasets. This is more work than downloading a pre-trained model, but the predictions will be reliable for your actual application.
Thermal history as a prediction feature
One consistently underweighted class of features in cycle life prediction models is thermal history. Most published models use temperature as a conditioning variable or a stratification factor but do not include cumulative thermal stress indicators as primary features. This leaves information on the table.
The argument for including thermal history directly is that the dominant degradation mechanisms (SEI growth, lithium plating risk) are temperature-dependent in ways that accumulate over the cell's lifetime. A cell that spends 30% of its cycles at 40 degrees Celsius and 70% at 25 degrees Celsius ages faster than one that spends all cycles at 25 degrees Celsius, and the differential is not fully captured by instantaneous temperature measurements. The cumulative heat dose, parameterized appropriately, is a better predictor of SEI growth extent than ambient temperature in any single cycle.
A practically useful thermal history feature is something like a weighted Arrhenius integral over the cell's thermal history: the integral of a degradation rate function (calibrated from the cell chemistry's activation energies) over time, weighted by temperature at each point. This can be accumulated by a BMS that records temperature at regular intervals and updated in real time. When used as a model input, it compresses the cell's full thermal history into a single number that is physically interpretable as a degradation accumulator.
Connecting prediction to protocol optimization
The purpose of cycle life prediction in the context we work in is not just diagnostic. It is prescriptive: given a predicted remaining useful life under the current operating protocol, what protocol modification would extend that life, and by how much?
This requires the model to be differentiable with respect to protocol parameters, or at least evaluable at multiple protocol scenarios efficiently enough to permit optimization. A pure black-box neural network trained on raw features can be accurate in prediction but opaque with respect to protocol sensitivity. A model with physically-grounded features that encode specific degradation mechanisms is more interpretable: you can ask which feature (SEI growth rate, plating extent, cathode stress) is the primary driver of the predicted end-of-life, and you can compute which protocol parameter change (lower charge cutoff voltage, reduced peak temperature, slower charge rate in high-SoC region) has the largest effect on that specific mechanism.
The payoff from this approach is that cycle life prediction feeds back into protocol design. The two are not separate exercises. You predict cycle life under the current protocol, identify the dominant degradation driver, modify the protocol to reduce that driver's rate, re-predict under the modified protocol, and iterate. This is the loop that connects the electrochemical understanding in the model to the engineering decision about how to charge the pack.
Whether machine learning adds value in this loop depends on what problem you are solving. If you have high-quality mechanistic models from first principles, pure ML may be redundant. If you have empirical data from real field cycles but limited physical characterization data, ML as a feature-augmented interpolation across your empirical data is the right tool. Most practical cases are somewhere between those extremes, and hybrid physics-informed approaches that combine mechanistic model structure with data-fitted parameters tend to outperform both pure approaches on generalization outside the training distribution. That is where the work is most interesting.