PhysicsNeMo-Mesh#

PhysicsNeMo-Mesh is a GPU-accelerated, PyTorch-based mesh module for physics-AI workflows. It handles point clouds, curves, surfaces, and volumes through a single API, integrates with PyTorch autograd, and provides discrete calculus, differential geometry, subdivision, repair, spatial queries, topology, and I/O.

Meshes across point clouds, curves, surfaces, and volumes

The word “mesh” means different things to different communities: CFD and FEM engineers think of volume meshes; graphics programmers think of surface meshes; computer vision researchers think of point clouds; robotics engineers think of curves. PhysicsNeMo-Mesh handles all of them.

This User Guide covers concepts, the data model, the capability surface, and usage patterns. The API reference documents every public function and class.

Hands-on tutorials are at examples/minimal/mesh/ in the PhysicsNeMo repository. Questions: GitHub Discussions.

Install with:

pip install "nvidia-physicsnemo[cu13,mesh-extras]"

Chapters#

Concepts and Data Model

Dimensional generality, the five-field Mesh tensorclass, why arbitrary-rank tensor fields matter, and the simplicial restriction.

Concepts and Data Model

Using Mesh

The capability matrix, common patterns (creating, transforming, calculus, surgery, serialization), and the mesh / graph / point-cloud equivalence.

Using Mesh

Simulation Domains with DomainMesh

Representing a full PDE problem (interior + named boundaries + global metadata) as a single DomainMesh.

Simulation Domains with DomainMesh

Mesh in ML Pipelines

Pipeline inversion, batching, feature extraction, data augmentation, and torch.compile compatibility.

Mesh in ML Pipelines