Accelerated Discovery of Aerospace Materials via Graph Neural Networks and Multiscale Modeling

The extreme aerothermal conditions encountered in space vehicle re-entry and supersonic flight pose severe challenges for material survivability. Window materials, which must maintain high optical transparency in addition to high mechanical strength, thermal stability, and chemical resistance, pose particular challenges. Here, we demonstrate high-throughput discovery of new high-temperature aerospace window materials using a combination of machine learning-guided materials screening, first-principles materials theory, macroscale modeling with custom thermo-mechanical-optical models, and high-fidelity computational fluid dynamics (CFD) simulations. Our machine learning models, powered by a novel integrated graph neural network (I-GNN) architecture, enable accurate predictions of key thermo-mechanical properties, including thermal conductivity, coefficient of thermal expansion, melting temperature, and other properties. Importantly, by integration of materials physics with GNN models, we demonstrate accurate, temperature-dependent predictions of these properties. Use of these fundamental property predictions as inputs to high-fidelity CFD simulations enables, for the first time, rapid and accurate predictions of macroscale material behavior under realistic aerothermal flow conditions for many thousands of previously unexplored materials. With this approach, several new candidate materials have been identified with combinations of thermal, mechanical, and optical properties that may enable superior performance compared with conventional sapphire windows in the mid-wave and long-wave IR spectral ranges, such as AlReSi and CaZrN₂. Property predictions for these and other top candidate materials have been validated using first-principles calculations with density functional theory and density functional perturbation theory.