Surrogate-assisted parameter identification for multiphysics lithium-ion battery models: Identifiability, optimization fidelity, and predictive transferability

Authors:
  • DNDuy Thanh Nguyen
    Duy Thanh Nguyen
  • MNMinh T. N. Nguyen
    Minh T. N. Nguyen
  • HPHai Hung Pham
    Hai Hung Pham
  • PNPhi Long Nguyen
    Phi Long Nguyen
  • MLMinh Bau Luong
    Minh Bau Luong
  • AGAkhil Garg
    Akhil Garg
  • HNHong Minh Nguyen
    Hong Minh Nguyen
  • HNHuu Dai Nguyen
    Huu Dai Nguyen
  • HNHoang Khan Nguyen
    Hoang Khan Nguyen

Abstract
Accurate parameterization of physics-based lithium-ion battery models is essential for reliable prediction of internal battery states, but repeated multiphysics model evaluations during optimization make parameter identification computationally demanding. Data-driven surrogates offer a promising alternative, but high predictive accuracy on test data does not necessarily imply reliable optimization, because surrogate errors in the regions explored during parameter search can alter the ranking of candidate solutions. This study evaluates this issue for an electrochemical–thermal–aging model of a commercial prismatic cell by comparing responseseries reconstruction and direct error-metric prediction within a unified surrogate-assisted identification framework. The results reveal that response-series surrogates provided more reliable parameter identification than direct error prediction, while the identified solutions also revealed parameter non-uniqueness, with different parameter combinations producing similar measurable responses. Individual surrogate evaluations were accelerated by several orders of magnitude, but the end-to-end identification speedups were limited to 3.97-fold and approximately 5.9-fold for the electrochemical–thermal and aging stages, respectively. Validation outside the calibration domain showed reasonable predictive performance under low C-rate conditions, although larger discrepancies remained in the temperature response. Beyond approximately 600 cycles, the experimental degradation increasingly diverged from the SEI and lithium plating degradation model, indicating when the existing degradation description begins to lose physical adequacy and when additional aging mechanisms may need to be considered

This publication has 32 references indexed in Scilit: