How nvidia/physicsnemo works
PhysicsNeMo is a PyTorch framework for physics and scientific machine learning, offering reusable model, numerical, data, and training components. A representative workflow is to load scientific data, prepare it for a model, train or run inference, and evaluate predictions; active-learning orchestration adds iterative training and data-selection workflows. The graph also preserves mesh processing, diffusion, and distributed/domain-parallel capabilities. The README establishes the product capabilities, while the source index provides only partial implementation evidence, so uncertain connections are left dashed.
Main entry points: models Model Architectures · datapipes Scientific Data · mesh.py Mesh Representation · driver.py Active Learning
Components 24
Data and Training
- Scientific Data
physicsnemo/datapipes - Data Readers
physicsnemo/datapipes/readers/numpy.py - Reader Registry
physicsnemo/datapipes/registry.py - Data Transforms
physicsnemo/datapipes/transforms - Active Learning
physicsnemo/active_learning/driver.py - Learning Loop
physicsnemo/active_learning/loop.py
Models and Generative AI
- Model Architectures
physicsnemo/models - Diffusion
physicsnemo/diffusion - Parameter-Efficient Tuning
physicsnemo/experimental/peft/io.py - Evaluation Metrics
physicsnemo/metrics
Mesh Computing
- Mesh Representation
physicsnemo/mesh/mesh.py - Mesh Calculus
physicsnemo/mesh/calculus/measure.py - Mesh Adjacency
physicsnemo/mesh/neighbors/_adjacency.py - Mesh Repair
physicsnemo/mesh/repair/pipeline.py
Distributed Computing
- Distributed Manager
physicsnemo/distributed/manager.py - Process Groups
physicsnemo/distributed/config.py - ShardTensor
physicsnemo/domain_parallel/shard_tensor.py - Shard Specifications
physicsnemo/domain_parallel/_shard_tensor_spec.py
Framework Foundation
- Model Base Module
physicsnemo/core/module.py - Filesystem Utilities
physicsnemo/core/filesystem.py - Optional Dependencies
physicsnemo/core/version_check.py - Feature Warnings
physicsnemo/core/warnings.py - Torch Profiler
physicsnemo/utils/profiling/torch.py
Other components
- ML Engineer
Connections 20
- ML Engineer → Active Learning: configures
- Active Learning → Learning Loop: runs
physicsnemo/active_learning/driver.py - ML Engineer → Scientific Data: provides
- Scientific Data → Data Readers: loads through
physicsnemo/datapipes/_rng.py - Data Readers → Reader Registry: registers with
physicsnemo/datapipes/readers/numpy.py - Scientific Data → Data Transforms: can prepare
- Data Transforms → Model Architectures: feeds
- Model Architectures → Evaluation Metrics: evaluated by
- Model Architectures → Diffusion: may use
- Model Architectures → Parameter-Efficient Tuning: can be tuned with
- Scientific Data → Mesh Representation: represents mesh data
- Mesh Representation → Mesh Calculus: supports
physicsnemo/mesh/mesh.py - Mesh Representation → Mesh Adjacency: supports
physicsnemo/mesh/mesh.py - Mesh Representation → Mesh Repair: can be repaired by
physicsnemo/mesh/repair/pipeline.py - Model Architectures → Distributed Manager: can scale with
README.md - Distributed Manager → Process Groups: uses configuration
physicsnemo/distributed/manager.py - ShardTensor → Shard Specifications: uses specifications
physicsnemo/domain_parallel/shard_tensor.py - Model Base Module → Filesystem Utilities: uses
physicsnemo/core/module.py - Model Base Module → Optional Dependencies: checks requirements
physicsnemo/core/module.py - Active Learning → Torch Profiler: references
physicsnemo/active_learning/driver.py