PhysicsNeMo: Neural-Operator-Based Physics Simulation Framework
Paper summary
Title: PhysicsNeMo: A Framework for Neural Operator-based Physics-ML Models Authors: NVIDIA Physics Simulation & AI team
Contributions
- Unified neural-operator framework (FNO / DeepONet / AFNO / Kapoor Transformer)
- End-to-end differentiable pipeline (geometry -> BC -> solve -> post)
- Multi-GPU/TPU distributed training (Megatron-style)
- Production deployment (ONNX Export + Triton)
Key results
- FNO accuracy 30% higher than traditional RANS surrogate on NASA Clap benchmark
- 8-GPU linear speedup ~= 7.5x
- End-to-end inference < 50 ms/sample (A100)
Use cases
- Industrial physics-AI products (auto, aerospace, energy)
- Large physics foundation-model pretraining
- Integration with NVIDIA Modulus (now PhysicsNeMo)
Notes
- Apache-2.0, commercial-friendly
- Needs CUDA 12+, A100/H100 recommended
Tags: NVIDIANeural OperatorFrameworkIndustrialMulti-GPU
Related entries
- Physics-Informed Learning Annual Review 2025-2026 · A systematic review of 2025-2026 progress in physics-informed ML - neural operat
- DeepONet (Original Paper) · Presents a branch-trunk architecture realizing a deep learning version of the un
- Fourier Neural Operator (Original Paper) · Introduces neural operators with integral kernels parameterized in Fourier space
- Transolver: A Transformer Solver for Large-Scale Industrial Geometries · Proposes a physics-attention mechanism that projects meshes onto learnable physi
- The Well: A Multi-Domain Benchmark for Physics Foundation Models · Proposes a unified benchmark across 16 physical domains, arguing for cross-domai