PHYSICS-CONSTRAINED VERIFICATION AND TELEMETRY BENCHMARK SUITE FOR DATA-DRIVEN ELECTRO-THERMAL AND CYBER-PHYSICAL SYSTEMS

Abstract

Data-driven predictive models and neural surrogates deployed across industrial Internet-of-Things (IoT) and cyber-physical hardware—such as lithium-ion battery management systems (BMS), high-density printed circuit board (PCB) power distribution networks, and liquid-cooled power electronics—are typically evaluated using aggregate statistical metrics like mean-squared error (MSE) on held-out test splits. However, standard statistical validation frequently masks catastrophic physical failure cliffs in out-of-distribution (OOD) operating regimes, where models achieve low average error on nominal trajectories yet violate fundamental conservation laws during transient overloads. This paper presents a deterministic, solver-in-the-loop verification harness and multi-domain electro-thermal telemetry benchmark comprising 9,600 operating scenarios across three cyber-physical subsystems: (1) 4S Li-Ion traction packs under pulse discharge audited for dT/dt exothermic thermal runaway, (2) high-current copper PCB traces audited against IPC-2152 electrothermal ampacity and return-path impedance rules, and (3) microchannel cold plates audited via coupled Jouleheating and Fourier heat-conduction finite-element balances. Experimental benchmarking demonstrates that while unconstrained neural surrogates miss 68.5% to 75.2% of physical thermal-runaway and ampacity violations under a standard MSE gate, coupling the evaluation harness with deterministic conservation-law and IPC-2152 solver checks detects 99.4% of physical failure cliffs and improves fault classification Macro F1 from 0.84 to 0.985.