NVIDIA PhysicsNeMo Release Notes

NVIDIA PhysicsNeMo v25.06 introduces significant improvements including a new DoMINO NIM with enhanced accuracy for vehicle aerodynamics, 10x faster training recipes, and 25x speedup for CorrDiff weather downscaling models. The release adds customizable validation benchmarks for physics-based AI model evaluation, new training samples for structural mechanics and flood dynamics modeling using graph neural networks, and reference workflows for hybrid solver initialization. Previous versions have progressively added support for various neural architectures like GraphCast, StormCast, and FourCastNet, along with distributed training capabilities and physics-informed neural operators for applications ranging from weather prediction to molecular dynamics.