Liquid Neural Networks for Battery Prognostics: Multi-Metric Evaluation and Interpretability Analysis

Abstract

Data-driven battery prognostics models are increasingly expected to be accurate, efficient, robust, and interpretable, yet existing benchmarks often report accuracy alone using random train/test splits that permit temporal leakage. This study presents a comprehensive multi-metric evaluation of liquid neural networks (LNNs) and neural circuit policies (NCPs) for capacity-based state-of-health prediction on the four-unit NASA Prognostics Center of Excellence (PCoE) lithium-ion battery dataset under a chronological splitting protocol designed to prevent future-cycle information leakage. On the shared raw-capacity benchmark, compact continuous-time architectures (under 10K parameters) obtain accuracy comparable to selected larger baselines with approximately 5-550x more parameters in the evaluated same-protocol and contextual comparisons; the main LNN-TS configuration trains 4-20x faster than conventional raw time-series baselines. An accuracy-robustness trade-off emerges across model families, with the most accurate models exhibiting 4.6-5.5% noise degradation and NCP sparsity providing a controllable trade-off between prediction quality and noise insensitivity. However, intrinsic uncertainty estimation through tau variability and hidden-state stability yields poorly calibrated confidence signals, identifying reliable self-assessment of prediction quality as an open challenge. Complementary interpretability experiments (confidence estimation, hidden-state trajectory analysis, phase-wise position importance, and counterfactual perturbation) suggest that LNNs learn representations correlated with physically plausible degradation behavior, while fixed-sparsity NCP experiments characterize accuracy-interpretability trade-offs without claiming learned causal pathways. Together, these results support compact continuous-time architectures as efficient and interpretable candidates for battery prognostics benchmarks that jointly evaluate accuracy, robustness, and model transparency.