Physics‐Informed Machine Learning for Solid Oxide Cell Materials and Microstructures Under Operando Constraints

Abstract

Solid oxide cells (SOCs) operate at high temperature under dual atmospheres and electrochemical polarization. Their performance depends not only on composition and instantaneous operating conditions, but also on operating history, because defect populations, transport pathways, and microstructures evolve during service. This behavior limits conventional machine‐learning models trained on static or steady‐state data, because high interpolation accuracy rarely guarantees reliable extrapolation to operando or aging conditions. In this review, physics‐informed machine learning (PIML) encompasses physical knowledge introduced through descriptors, model architectures, generative constraints, and experimental feedback, rather than only neural networks constrained by partial differential equations. We organize representative SOC studies using a four‐level hierarchy of physics embedding and relate it to data fidelity, while tracing descriptors from elemental attributes to dynamic operando‐state variables. We then examine how connectivity, conservation laws, and thermodynamic stability can constrain microstructure inverse design and reject physically inadmissible outputs. Further progress requires closer links among defect chemistry, microstructure physics, service‐relevant data, and experimental feedback, enabling models that remain valid under realistic conditions and support closed‐loop design.