This project will develop an AI-driven multiscale workflow for transferring atomistic knowledge from universal machine-learning interatomic potentials (uMLIPs) to mesoscale phase-field simulations andrepresentative atomistic mechanical validation.
The target systems are Cu FCC, W BCC and Ti HCP, covering three representative crystal structures. uMLIPs will be used as DFT-informed knowledge sources for seed-data generation, uncertainty estimation and uncertainty-aware distillation, rather than as direct production simulation engines. Their outputs will be distilled into efficient system-specific UF3 potentials, which will enable active learningdriven atomistic calculations of grain-boundary energy and mobility. Machine-learning surrogate models will then convert the resulting discrete atomistic data into continuous grain-boundary property functions for GPU-accelerated phase-field simulations. Representative phase-field microstructures will be reconstructed into atomistic models and validated by MD nanoindentation simulations.
The requested EuroHPC resources are required for the repeated workflow cycles of uMLIP reference, UF3 distillation, grain-boundary calculations, surrogate-model training, GPU phase-field calibration and atomistic validation. The expected outcome is a reusable AI-for-Science workflow for atomistically informed microstructure evolution modelling.
Principal Investigator, Company and Country
Kai Liu, Luxembourg Institute of Science and Technology, Luxembourg