NVIDIA PhysicsNeMo Physics-AI FrameworkNew

What it is

PhysicsNeMo (formerly NVIDIA Modulus) is an open-source framework for building, training and deploying physics-ML models. Compared with research libraries, it focuses on engineering concerns: distributed training, mixed precision, geometry parameterization, and inference export.

Key modules

• Domain solvers: PINN / FNO training recipes with ready-made PDE losses
• Geometry: parameterized constructive geometry (NVIDIA Warp based) for shape optimization loops
• Performance: multi-GPU, TF32/BF16 mixed precision, torch.compile friendly
• Inference: ONNX / TensorRT export paths for production serving

Learning path

1. Start from examples/cfd - reproduce the laminar cylinder flow tutorial
2. Move to examples/structural_mechanics - FNO for displacement fields
3. Study examples/additive_manufacturing - melt pool temperature prediction
4. Explore examples/tcad - semiconductor device simulation surrogates

When to choose it

• You need multi-GPU training or fast inference on industrial meshes
• You want reference implementations that mirror CAE vendor workflows
• Trade-off: heavier dependency stack than minimal research libraries