Endpoint Fidelity and Chemical Transfer in Machine Learning for Lithium-Ion Cathode Voltage Prediction

Authors

  • Chao Gao Associate Professor, Department of ICT and Natural Sciences, Norwegian University of Science and Technology (NTU), Ålesund, Norway Author
  • Anniken Karlsen Associate Professor, Department of ICT and Natural Sciences, Norwegian University of Science and Technology (NTU), Ålesund, Norway Author

Keywords:

Lithium-Ion Cathodes, Voltage Prediction, Endpoint Composition, Nested Validation, Chemical Transfer, Conformal Prediction

Abstract

Composition-based voltage models offer inexpensive screening of lithium-ion electrode materials, but a completely Li-removed formula may misrepresent the charged endpoint associated with the target voltage. This study tests the value and deployment limits of endpoint information using 2,636 Materials Project electrode records spanning 1,763 discharged compositions and 601 chemical systems. Pointwise composition descriptors were regenerated, and discharged-only, Li-removed-proxy, and actual-endpoint representations were compared using five outer chemical-system folds with three-fold inner parameter selection. All fitted preprocessing remained within training partitions. The recorded charged endpoint retained Li in 882 records. Under chemical-system holdout, actual endpoints yielded mean absolute error 0.455 V, root mean squared error 0.619 V, and R^2 0.654, compared with 0.470 V, 0.642 V, and 0.628 for the proxy representation. The paired RMSE difference was -0.0234 V, with a conditional chemical-system bootstrap interval of [-0.0391, -0.0094] V. A secondary equal-dimensional endpoint control yielded RMSE 0.621 V. Composition-grouped evaluation produced actual-endpoint RMSE 0.573 V. Split-conformal intervals reached 96.2% empirical coverage at a nominal 95% level, with mean full width 2.861 V. Separately, canonical identity auditing reduced an archived 1,056-record filtered candidate set to 702 training-excluded records. These findings support a modest endpoint-information benefit while distinguishing reaction-informed prediction, formula-only screening, and independent materials verification.

Downloads

Published

2026-03-30

Issue

Section

Articles

How to Cite

Endpoint Fidelity and Chemical Transfer in Machine Learning for Lithium-Ion Cathode Voltage Prediction. (2026). International Journal of Engineering Innovation and Computational Science, 1(1), 50-61. https://journovapublishing.com/journals/ijeics/article/view/7

Similar Articles

You may also start an advanced similarity search for this article.