A lithium-ion battery degrades gradually until it doesn't. The capacity fade is nearly linear for hundreds of cycles, then accelerates sharply — the degradation knee. After the knee, the cell fails fast. Predicting when the knee arrives is the difference between a battery that lasts and one that doesn't.
A world model for battery degradation (arXiv:2603.10527) encodes raw voltage, current, and temperature time-series from each charge cycle into a latent state, then propagates that state forward through learned dynamics. The critical finding: iterative rollout — stepping the dynamics model forward cycle by cycle — halves the trajectory forecast error compared to direct regression from the same encoder. The model that simulates the future outperforms the model that jumps to it.
Note (2026-11-14; read against the paper's full text): the essay says the physics enters through a Single Particle Model constraint in the training loss, and that without it the model misses the nonlinearity. The arXiv abstract page says a Single Particle Model constraint, but the paper's own text and abstract describe a monotonicity penalty derived from irreversible degradation, that is, a non-increasing state-of-health trajectory; the Single Particle Model appears in its introduction as background. It says this penalty is most effective at the knee where capacity fade accelerates and the model is prone to predicting temporary recovery, that it improves prediction at the knee without changing aggregate accuracy, and that below a state of health of 0.90 the constraint over-regularizes. The experiments use the Severson LiFePO4 dataset of 138 cells. I found no discussion of SEI growth or lithium plating in the text.
The physics enters through a Single Particle Model constraint in the training loss. The SPM connects internal resistance changes to capacity fade — a relationship that's most predictive exactly at the degradation knee, where resistance-driven failure accelerates. Without the constraint, the learned dynamics fit a smooth trajectory that misses the nonlinearity. With it, the model captures mechanism-specific behavior precisely when generic fitting fails.
The structural insight: the knee isn't an anomaly in the degradation trajectory. It's where one degradation mechanism (e.g., SEI growth) hands off to another (e.g., lithium plating). A purely data-driven model sees a curve. A physics-constrained model sees a regime change. The constraint doesn't just improve accuracy — it changes what the model represents. Fitting a curve is interpolation. Modeling a regime change is understanding.