PIGNet V2: Physics-Informed Graph Neural Networks for High-Throughput Crystalline Material Property Prediction

PIGNet V2 is a physics-informed graph neural network framework for high-throughput crystalline materials property prediction. The model combines attention-gated message passing, 56-dimensional 3-body angular edge featurisation, thermodynamically constrained multi-task learning, and conformal uncertainty calibration to predict band gap, formation energy, and energy above hull directly from crystal structures. PIGNet V2 introduces physics-constrained Softplus output heads that guarantee non-negative physically valid predictions by construction, while the companion BatteryFormer architecture enables inference on unrelaxed crystal structures for accelerated materials screening workflows. The framework is trained on 125,000 Materials Project structures and is designed for computational materials discovery, battery cathode optimisation, and scientific machine learning applications. This upload contains the preprint manuscript associated with the PIGNet V2 framework developed by Scandium Labs Research Group. GitHub repository:https://github.com/shamiquekhan/Scandium-Lab-Model