Companion packages

Four things ship alongside the core library. Two are inside the physicsnemo package; two are separate installs.

physicsnemo.active_learning (bundled)

The loop that lets the model choose its own training solves. It has its own page: Active learning.

physicsnemo.experimental (bundled)

Not stable API, but several things here are the only implementation of what they do:

Module	What it holds
experimental.guardrails.embedded	OODGuard, OODGuardConfig — out-of-distribution detection on a model's inputs
experimental.guardrails.geometry	GeometryGuardrail, extract_features, validate_mesh — out-of-distribution detection on the shape of a triangular surface mesh (2.2)
experimental.uq	variational GP heads (variational_gp_head, field_variational_gp_head) for calibrated posterior variance
experimental.peft	LoRA adapters — apply, merge, io
experimental.models	flare, geotransolver, aerojepa, globe, healda, strata, xdeeponet, diffusion, diffusion_unets
experimental.nn	FLARE attention, point tokenizers, RoPE, symmetry layers, 3-D diffusion U-Net blocks
experimental.datapipes, experimental.metrics, experimental.utils	HealDA pipeline, diffusion metrics, caching and prefetch

One note for anyone doing volumetric diffusion. physicsnemo.models.diffusion_unets is 2-D-image oriented, but experimental.models.diffusion_unets.DiffusionUNet3D is a genuine volumetric denoiser: it implements the physicsnemo.diffusion.base.DiffusionModel protocol, so it composes with the same preconditioners, losses and samplers, and takes optional volume ((B, C, D, H, W)) and vector conditioning.

physicsnemo-cfd (separate install, source only)

pip install git+https://github.com/NVIDIA/physicsnemo-cfd

Not on PyPI. Provides physicsnemo.cfd:

• cfd.postprocessing_tools.metric_registry — a domain-aware CFD metric registry (relative-L2, drag and lift, physics residuals, UQ metrics);
• cfd.hybrid_initialization_tools — blending a prediction into a solver's initial condition;
• cfd.evaluation — checkpoint-driven evaluation wrappers, benchmarks, datasets, reports and NIM clients. In the 0.0.3a0 release the benchmark engine (run_benchmark, write_report) is declared but does not import.

physicsnemo-curator (separate install, git only)

An ETL framework for turning raw simulation output into AI-ready datasets (Zarr stores, VTU grids). A pipeline is Source -> Filter -> Sink, run sequentially or over a process pool; its sinks ship upstream, and the source side is what a solver has to supply.

Its build pulls a Rust toolchain, which is a real installation cost for what it does. Since 2.2 the mesh package's own to_zarr writes comparable AI-ready Zarr without it, and — for the columnar case — so does write_dataset, in pure Python over pyarrow or zarr.

Optional dependencies at a glance

Package	Needed for	Without it
torch	everything ML	meshio++ still imports; ML entry points raise with an install hint
nvidia-physicsnemo	everything ML	same
torch_geometric, torch_scatter	the graph models and their batching	those paths raise
torch_sparse or dgl	GraphCast	that recipe is unavailable
onnxruntime / onnxruntime-gpu	ONNX inference	ONNX paths raise
tritonclient	Triton serving	Triton paths raise
gpytorch	the GP uncertainty head	that head raises
pyvista	mesh visualization	those paths raise
nvidia-physicsnemo-cfd	the CFD metric registry	its metrics are unavailable
physicsnemo-curator	its Zarr/VTU AI-ready export	that pipeline is unavailable
onnxscript	torch's ONNX exporter	export raises with an install hint
cupy	GPU array interchange	falls back to numpy
usd-core	OpenUSD export	that path raises (meshio++'s own usd format needs it too, via the [usd] extra)
tetgen	exact boundary recovery in a tetrahedral fill	that backend raises; the default is unaffected

In meshio++

The same lazy-gate policy applies here, and it is enforced rather than intended: import meshioplusplus succeeds with none of the above installed, and the default CI matrix runs the whole test suite that way. Every gated surface raises a named error saying what to install.

Two of those messages are deliberately not pip install meshioplusplus[...], and the reason is the same in both cases — the wheel is not pinnable for you. torch's default Linux wheel bundles ~900 MB of CUDA, nvidia-physicsnemo hard-depends on a particular torch, and CuPy ships one wheel per CUDA major (cupy-cuda13x, cupy-cuda12x, ROCm). An extra that resolved to the wrong one would be worse than no extra, so those errors name the package and let you pick the build.

The extras that do exist are the ones with a single correct answer: [arrow], [pandas], [polars], [zarr], [pyvista], [trimesh], [interop], [viewer], [mcp], [kahip], [codecs], and [all] for the optional dependencies the formats need. See Installation.

Next: Versions and compatibility.