Human-Guided AI Development of Physics-Bounded Screening Rules for Sparse Battery Signals: A LiFePO4 Case Study

Abstract

Battery-management systems observe current, terminal voltage, limited temperature measurements, and operating setpoints, but not the internal variables of electrochemical theory, so routine signals generally cannot identify a unique mechanism. We present a human-guided, generative-AI-assisted methodology that translates physics-based expectations into deterministic, auditable screening rules: human scientific authority fixes the physical assumptions, evidence requirements, and permissible claims, artificial intelligence supports development, and runtime evaluation is non-generative. LiFePO4 is the test case. A Zeng–Bazant current-dependent plateau approximation supplies a physics-based reference, and a bivariate representational precedent motivates a composite, reference-dependent voltage residual that is not identified as thermodynamic work. Eleven observable screens return present, absent within resolution, indeterminate, or unavailable. Four evidence forms are separated. Digitized model curves show the reduced plateau relation tracks its parent phase-field simulation through moderate rates, with a high-rate limitation. Published temperature-conditioned discharge profiles show stable plateau elevation and flattening from 268 to 298 K across 0.5C–2C, with mixed 2C curvature. Measured replicates of a commercial cylindrical cell at two ambient setpoints resolve a within-run surface-temperature depression whose integrated first-law balance is heat-rejection-dominant and compatible with, but not uniquely attributed to, a literature-bounded reversible contribution. Controlled synthetic cases verify deterministic feature recovery and abstention without establishing a mechanism.