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

Functional oxide materials are widely employed across a range of electronic applications, where electrical conductivity and its temperature dependence serve as key screening metrics for material suitability. Across these applications, experimental characterisation remains resource-intensive and time-consuming. As a result, machine learning approaches have recently emerged as a digital alternative to accelerate materials discovery processes. However, datadriven approaches are typically restricted by the lack of a comprehensive experimentally-sourced database, with existing models trained on simulated density-functional theory (DFT) calculations. Here, we compiled an experimental database of over 7,000 temperature-dependent electrical conductivity samples for oxides, primarily extracted from commercial materials databases and peer-reviewed literature. We benchmark MODNet, a state-of-the-art general purpose composition-based machine learning baseline, against TriCondNet (a three-specialist conductivity network), a novel physics-guided neural network (PGNN) architecture developed to capture temperature-dependent electrical conductivity behaviour of metallic and semiconducting oxides. Our results demonstrate that TriCondNet outperforms baseline models on standard error metrics while adhering to the physical models for temperature–conductivity dependence. We then applied our top-performing model to an extensive composition list, screening candidates for two different application-specific case studies: dielectric and thermoelectric materials. Candidates were ranked based on their required conductivity relevant to the case study and associated uncertainty estimates. The resulting top ranked candidates matched materials already recognised as leading dielectrics and thermoelectrics, confirming that the workflow recovers established high-performers and establishing TriCondNet as a reliable basis for identifying new candidates through composition interpolation across known oxide families.