What this research found
Can a lithium-ion cell's total lifetime be forecast from its first 100 cycles, before any measurable capacity fade? Working from the public 124-cell fast-charging benchmark of Severson et al. (2019), K-Dense engineered 17 early-cycle features under a code-enforced guard that blocks any read past cycle 100, then compared a regularized linear model against a random forest. The Elastic Net predicted the 40-cell held-out batch with a root-mean-square error of 189.4 cycles and a mean absolute percentage error of 11.82%, while the higher-capacity forest generalized worse.
- A single early-cycle statistic carries most of the signal: the base-10 logarithm of the variance of the cycle-100-minus-cycle-10 discharge-capacity difference curve correlates with log cycle life at Pearson r = -0.927 across all 124 cells. No feature outside that curve's moments exceeds an absolute r of 0.68.
- The multi-feature Elastic Net reached a Secondary-Test root-mean-square error of 189.4 cycles and a mean absolute percentage error of 11.82%, with a median absolute percentage error of 8.8%. Its errors grew only modestly from training (84.1 cycles) through the primary test batch (90.8 cycles).
- The 500-tree random forest had the lowest training error (7.07% mean absolute percentage error) but the worst held-out error (292.2 cycles, 14.88%), reproducing the original study's preference for a sparse regularized linear predictor.
- On the held-out batch, 22 of 40 cells were predicted within 10% of their true life and 27 of 40 within 15%, with a predicted-versus-observed correlation of r = 0.79.
- Errors concentrate in the longest-lived cells, which the model systematically under-predicts: 1,935 observed versus 1,168 predicted for one cell, 1,836 versus 1,380 for another, and 1,801 versus 1,398 for a third.
How it was done
The three raw batch files from the Severson dataset were cleaned following the authors' protocol, leaving 124 commercial LFP/graphite 18650 cells with observed lives spanning 148 to 2,237 cycles. Features were drawn from four physically motivated groups computed strictly over cycles 2 to 100: moments of the discharge-capacity difference curve, the early capacity trajectory, the average-temperature stream, and internal resistance. A runtime leakage guard raised an error on any access beyond cycle 100, and a per-cell audit confirmed the maximum cycle touched was exactly 100 for all cells. Models were trained on the base-10 logarithm of cycle life using the published batch division: 41 training cells, 43 primary-test cells for hyperparameter selection, and 40 chronologically later cells as a never-touched held-out batch.
Data sources
- Severson et al., Nature Energy 4:383-391 (2019) — 124 LFP/graphite cells, 72 fast-charging protocols, cycle lives 148-2,237
- MIT-Stanford-Toyota battery data platform — per-cycle discharge-capacity curves on a 1,000-point voltage grid
- Attia et al., Nature 578:397-402 (2020) — closed-loop fast-charging protocol optimization
- Roman et al., Nature Machine Intelligence 3:447-456 (2021) — machine-learning pipeline for state-of-health estimation
Limitations
The data covers a single cell chemistry and form factor cycled under controlled laboratory conditions, so transfer to other chemistries or noisy field data is untested. The model is a point predictor with no uncertainty quantification, and its feature set appears to saturate for the healthiest cells, producing a persistent under-prediction of the longest-lived ones.
How this research was produced
K-Dense Web planned and ran this materials investigation end to end — gathering the sources, carrying out the analysis, producing the figures, and drafting the report. The full session transcript, including every intermediate step, is available to view.
