Physics‐Guided Modeling of Temperature‐Dependent Electrical Conductivity of Functional Oxides

Functional oxides are widely used in electronic applications, where the temperature dependance of electrical conductivity is a key metric for material suitability. Experimental characterization of these properties remains resource‐intensive and time‐consuming, so machine learning has emerged as an alternative to accelerate materials discovery processes. However, data‐driven approaches are restricted by the lack of a comprehensive experimentally sourced database, with existing models trained on simulated density‐functional theory (DFT) calculations. Here, an experimental database of approximately 8000 temperature‐dependent conductivity samples for oxides is compiled from commercial databases and peer‐reviewed literature. MODNet, a state‐of‐the‐art general purpose composition‐based baseline, is benchmarked against TriCondNet, a physics‐guided neural network (PGNN) developed to capture temperature‐dependent conductivity behavior of metallic and semiconducting oxides. TriCondNet outperforms baseline models on standard error metrics while adhering to the physical models for conductivity–temperature dependance. The top‐performing model is then applied to an extensive composition list, screening candidates for two application‐specific case studies: dielectric and thermoelectric materials. Ranked by their predicted conductivity and associated uncertainty, the top candidates match materials already recognized as leading dielectrics and thermoelectrics. These results indicate that the workflow can recover established high‐performers, supporting the use of TriCondNet as a basis for identifying new oxide candidates through composition interpolation.