Merge branch 'main' into update-nightly-uv-pipeline

Author: coreyjadams

.github/CODEOWNERS

@@ -113,6 +113,9 @@ physicsnemo/models/figconvnet @coreyjadams
 # FNO
 # NEEDSOWNER
 
+# GLOBE
+physicsnemo/models/globe @peterdsharpe
+
 # GNN layers, mesh graphnet, graphcast
 physicsnemo/models/gnn_layers @mnabian
 physicsnemo/models/meshgraphnet @mnabian
@@ -143,6 +146,9 @@ physicsnemo/models/vfgn @mnabian
 
 # EXAMPLE SPECIFIC OWNERSHIP
 
+# GLOBE
+examples/cfd/external_aerodynamics/globe @peterdsharpe
+
 # corrdiff
 examples/weather/corrdiff @CharlelieLrt
 examples/active_learning @laserkelvin @dallasfoster
@@ -156,5 +162,8 @@ test/utils/generative @CharlelieLrt
 test/utils/corrdiff @CharlelieLrt
 test/active_learning @laserkelvin @dallasfoster
 
+# GLOBE
+test/models/globe @peterdsharpe
+
 # Mesh
 test/mesh @peterdsharpe

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -31,7 +31,7 @@ body:
     attributes:
       label: Version
       description: What version of PhysicsNeMo are you running?
-      placeholder: "example: 1.3.0"
+      placeholder: "example: 2.0.0"
     validations:
       required: true
 

.gitignore

@@ -165,7 +165,10 @@ cython_debug/
 nsight-systems*
 build/
 mlruns/
+mlflow.db
 checkpoints/
+cache/
+output/
 
 # Hydra
 outputs/

.importlinter

@@ -88,7 +88,7 @@ containers=
    physicsnemo.models
 layers = 
    mesh_reduced
-   afno | dlwp | dlwp_healpix | dit |domino | dpot | fengwu | figconvnet | fno | graphcast | meshgraphnet | pangu | pix2pix | rnn |  srrn | swinvrnn | topodiff | transolver  | vfgn | diffusion_unets
+   afno | dlwp | dlwp_healpix | dit | domino | dpot | fengwu | figconvnet | fno | graphcast | meshgraphnet | pangu | pix2pix | rnn |  srrn | swinvrnn | topodiff | transolver  | vfgn | diffusion_unets
    unet
 
 [importlinter:contract:physicsnemo-diffusion]
@@ -99,7 +99,8 @@ containers=
 layers = 
    generate
    samplers : metrics
-   noise_schedulers | multi_diffusion | preconditioners | guidance
+   multi_diffusion : preconditioners : guidance
+   noise_schedulers
    utils
 
 [importlinter:contract:physicsnemo-external-imports]

.pre-commit-config.yaml

@@ -46,6 +46,7 @@ repos:
       name: license
       entry: python test/ci_tests/header_check.py
       language: python
+      additional_dependencies: ['GitPython']
       files: \.(py|yaml|ci|release)$|Dockerfile$
       exclude: ^(physicsnemo/internal/|docs/)
 

CHANGELOG.md

@@ -6,7 +6,41 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
-## [2.0.0a0] - 2026-XX-YY
+## [2.1.0a0] - 2026-XX-YY
+
+### Added
+
+- Adds GLOBE model (`physicsnemo.experimental.models.globe.model.GLOBE`)
+- Adds GLOBE AirFRANS example case (`examples/cfd/external_aerodynamics/globe/airfrans`)
+- PhysicsNeMo-Mesh now supports conversion from PyVista/VTK/VTU meshes that may
+  contain polyhedral cells.
+- In PhysicsNeMo-Mesh, adds `Mesh.to_point_cloud()`, `.to_edge_graph()`, and
+  `.to_dual_graph()` methods. These allow Mesh conversion to 0D point clouds, 1D
+  edge graphs, and 1D dual graphs, respectively, when connectivity information
+  is not needed.
+
+### Changed
+
+- Improved crash recipe with configurable stats directory.
+
+### Deprecated
+
+- `physicsnemo.utils.mesh` is deprecated and will be removed in v2.2.0. For
+  isosurface extraction, use `physicsnemo.mesh.generate.marching_cubes` instead
+  of `sdf_to_stl`. For VTP/OBJ/STL file conversion (`combine_vtp_files`,
+  `convert_tesselated_files_in_directory`), use VTK or PyVista directly.
+
+### Removed
+
+### Fixed
+
+- Fixed bug in Pangu, FengWu attention window shift for asymmetric longitudes
+
+### Security
+
+### Dependencies
+
+## [2.0.0] - 2026-03-09
 
 ### Added
 
@@ -27,16 +61,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - DiT (Diffusion Transformer) has been moved from `physicsnemo.experimental.models.dit`
   to `physicsnemo.models.dit`.
 
-### Deprecated
-
-### Removed
-
 ### Fixed
 
 - Shape mistmatch bug in the Lennard Jones example
 
-### Security
-
 ### Dependencies
 
 - CUDA backend is now selected via orthogonal `cu12` / `cu13` extras rather

CONTRIBUTING.md

@@ -25,6 +25,68 @@ we will get back to you.
 
 ## Contribute to PhysicsNeMo-Core
 
+### Codebase Structure
+
+- [physicsnemo](physicsnemo) The core python components of the framework live
+here.  They are listed below not alphabetically but in hierarchical order
+(foundational layers first)
+  - [core](physicsnemo/core) Core utilities for `physicsnemo`, including
+  tools for model registration, filesystems, warnings, and optional imports.
+  - [distributed](physicsnemo/distributed) Basic distributed utilities to
+  enable data parallel training, and some advanced functionality such as
+  distributed FFTs and more sophisticated functional calls with autograd enabled.
+  - [utils](physicsnemo/utils) Utility tools common across physicsnemo.
+  Includes profiling tools, logging tools, model checkpointing and restoration
+  tools.  Note that model-specific tools should *not* go here.  Several
+  classes here are slowly being migrated to other locations or deprecated.
+  - [nn](physicsnemo/nn) `physicsnemo` offers a number of GPU-optimized,
+  standalone layers and functions through `physicsnemo.nn` and
+  `physicsnemo.nn.functional`, with a similar interface to `torch.nn`.
+  These layers are meant to be compatible with `torch` modules and you
+  should feel free to mix and match them into your models.  We offer,
+  for several functional calls, highly optimized and custom backends with
+  significant speedups over pytorch - see our documentation to learn more.
+  - [models](physicsnemo/models) Our model repository contains a set of
+  popular scientific AI models implemented with performance enhancements or
+  other features not easily available elsewhere. Many models are also
+  featured in training recipes in [examples](examples).
+  - [datapipes](physicsnemo/datapipes) offers GPU-accelerated datapipes for
+  scientific AI datasets.  We offer both bespoke datapipes for specific
+  problems, as well as a generic, GPU-focused data pipe abstraction similar
+  to `torchvision`.  This is designed to enable GPU preprocessing of high
+  resolution scientific data needed for AI for Science.
+  - [metrics](physicsnemo/metrics) A collection of general as well as domain
+  specific metrics for AI for Science model training and inference.
+  - [domain_parallel](physicsnemo/domain_parallel) Unique to PhysicsNeMo, the
+  `ShardTensor` abstraction provides a generic interface for domain
+  parallelism enabling extremely high resolution data training on batch size
+  per GPU < 1.  Read our
+  [documentation](https://docs.nvidia.com/physicsnemo/latest/user-guide/domain_parallelism_entry_point.html)
+  to find out more.
+  - [active_learning](physicsnemo/active_learning) The Active Learning module
+  provides scaffolding for building automated and end-to-end workflows using
+  active learning techniques.
+  - [diffusion](physicsnemo/diffusion) `physicsnemo` diffusion offers a
+  powerful, yet abstracted, collection of diffusion tools to build, train,
+  and inference a diffusion workflow for scientific problems.
+  - [experimental](physicsnemo/experimental) New functionality that may not
+  yet be API-stable is introduced through the experimental layer.  The
+  experimental folder mimics the physicsnemo structure under it's own
+  namespace, though only in places actively under development.  Models
+  especially are always introduced through experimental.
+- [test](test) `physicsnemo` tests live here.  New functionality in the core
+repository must include test coverage here.  Example code can not be tested
+from this folder but must be tested in examples.
+- [benchmarks](benchmarks) We have automated benchmarking tools for several
+components in the library - it's not expected that external contributions will
+need to know what's in this folder, though you're welcome to explore and
+investigate as you like.
+- [docs](docs) Our core documentation folder for the `physicsnemo` API.
+If you submit changes to the API, we welcome you to update the docs as well.
+- [examples](examples) We have a variety of examples to demonstrate the
+functionality of `physicsnemo` to explore.  Please visit the examples folder
+and README to learn more.
+
 ### Coding Style
 
 Beyond using standard tools for formatting and linting, we document and enforce

Dockerfile

@@ -89,24 +89,8 @@ RUN if [ "$TARGETPLATFORM" = "linux/amd64" ]; then \
     fi
 
 # Install vtk and pyvista
-ARG VTK_ARM64_WHEEL
-ENV VTK_ARM64_WHEEL=${VTK_ARM64_WHEEL:-unknown}
-
-RUN if [ "$TARGETPLATFORM" = "linux/arm64" ] && [ "$VTK_ARM64_WHEEL" != "unknown" ]; then \
-        echo "VTK wheel $VTK_ARM64_WHEEL for $TARGETPLATFORM exists, installing!" && \
-        uv pip install /physicsnemo/deps/${VTK_ARM64_WHEEL}; \
-    elif [ "$TARGETPLATFORM" = "linux/amd64" ]; then \
-        echo "Installing vtk for: $TARGETPLATFORM" && \
-        uv pip install "vtk>=9.2.6"; \
-    else \
-        echo "No custom wheel or wheel on PyPi found. Installing vtk for: $TARGETPLATFORM from source" && \
-        apt-get update && apt-get install -y libgl1-mesa-dev && \
-        git clone https://gitlab.kitware.com/vtk/vtk.git && cd vtk && git checkout tags/v9.4.1 && git submodule update --init --recursive && \
-        mkdir build && cd build && cmake -GNinja -DVTK_WHEEL_BUILD=ON -DVTK_WRAP_PYTHON=ON /workspace/vtk/ && ninja && \
-        python setup.py bdist_wheel && \
-        uv pip install dist/vtk-*.whl && \
-        cd ../../ && rm -r vtk; \
-    fi
+# VTK has started shipping aarch64 wheels, hence we no longer need a custom installation. Ref: https://pypi.org/project/vtk/9.6.0/#files
+RUN uv pip install "vtk>=9.6.0"
 RUN uv pip install "pyvista>=0.40.1"
 
 # Install onnxruntime (custom wheel for ARM)
@@ -223,13 +207,13 @@ RUN if [ "$TARGETPLATFORM" = "linux/amd64" ] && [ "$NATTEN_AMD64_WHEEL" != "unkn
 RUN uv pip install --no-build-isolation "torch_sparse"
 
 # All pyproject extras (no dev); installs physicsnemo non-editable
-RUN cd /physicsnemo && uv pip install ".[cu13,utils-extras,mesh-extras,datapipes-extras,gnns,perf]"
+RUN cd /physicsnemo && uv pip install ".[cu13,utils-extras,mesh-extras,datapipes-extras,gnns]"
 
 # Cleanup builder stage
 RUN rm -rf /physicsnemo/
 
 #######################################################################
-# CI image: builder + dev group + netcdf4 hack + FigNet/Makani + CI-only packages
+# CI image: builder + dev group + FigNet/Makani + CI-only packages
 #######################################################################
 FROM builder AS ci
 
@@ -240,8 +224,7 @@ ENV UV_SYSTEM_PYTHON=1
 ENV UV_BREAK_SYSTEM_PACKAGES=1
 ENV UV_CONSTRAINT=/etc/pip/constraint.txt
 
-# TODO: Remove hacky downgrade of netCDF4. netCDF4 v1.7.1 issue: https://github.com/Unidata/netcdf4-python/issues/1343
-RUN uv pip install "netcdf4>=1.6.3,<1.7.1"
+RUN uv pip install "netcdf4>1.7.3" dask
 
 COPY . /physicsnemo/
 
@@ -252,15 +235,15 @@ RUN cd /physicsnemo && uv pip install --group dev
 RUN FORCE_CUDA_EXTENSION=1 uv pip install --no-build-isolation "torch-harmonics==0.8.0"
 RUN uv pip install "tensorly>=0.8.1" "tensorly-torch>=0.4.0" "torchinfo>=1.8" "webdataset>=0.2"
 # Install Makani via direct URL
-RUN uv pip install --no-deps "git+https://github.com/NVIDIA/makani.git@v0.2.1#egg=makani"
+# RUN uv pip install --no-deps "git+https://github.com/NVIDIA/makani.git@v0.2.1#egg=makani"
 
 # Other CI-only specs (moto, scikit-image, etc.)
 RUN uv pip install "moto[s3]>=5.0.28"
 RUN uv pip install "numpy-stl" "scikit-image>=0.24.0" "sparse-dot-mkl" "shapely"
 RUN uv pip install "multi-storage-client[boto3]>=0.33.0"
 
 # E2Grid install
-RUN uv pip install --no-deps --no-build-isolation "git+https://github.com/NVlabs/earth2grid.git@11dcf1b0787a7eb6a8497a3a5a5e1fdcc31232d3"
+# RUN uv pip install --no-deps --no-build-isolation "git+https://github.com/NVlabs/earth2grid.git@11dcf1b0787a7eb6a8497a3a5a5e1fdcc31232d3"
 
 # Uninstall the non-editable physicsnemo from builder
 RUN uv pip uninstall nvidia-physicsnemo
@@ -301,6 +284,5 @@ ENV UV_BREAK_SYSTEM_PACKAGES=1
 ENV UV_CONSTRAINT=/etc/pip/constraint.txt
 
 # Install packages for Sphinx build
-RUN uv pip install "protobuf==3.20.3"
-RUN uv pip install "recommonmark==0.7.1" "sphinx==5.1.1" "sphinx-rtd-theme==1.0.0" "pydocstyle==6.1.1" "nbsphinx==0.8.9" "nbconvert==6.4.3" "jinja2==3.0.3"
+RUN uv pip install "recommonmark>=0.7.1" "sphinx>=5.1.1" "nvidia-sphinx-theme>=0.0.7" "pydocstyle>=6.1.1" "nbsphinx>=0.8.9" "nbconvert>=6.4.3" "jinja2>=3.0.3"
 RUN wget https://github.com/jgm/pandoc/releases/download/3.1.6.2/pandoc-3.1.6.2-1-amd64.deb && dpkg -i pandoc-3.1.6.2-1-amd64.deb

Makefile

@@ -29,14 +29,14 @@ lint:
 	pre-commit run markdownlint -a && \
 	pre-commit run check-added-large-files -a
 
-license: 
-	python test/ci_tests/header_check.py --all-files
+license:
+	pre-commit run license -a
 
 doctest:
 	coverage run \
 		--rcfile='test/coverage.docstring.rc' \
 		-m pytest \
-		--doctest-modules physicsnemo/ --ignore-glob=*internal* --ignore-glob=*experimental*
+		--doctest-modules physicsnemo/ --ignore-glob=*internal* --ignore-glob=*experimental* --ignore-glob=*deploy/onnx*
 
 pytest: 
 	coverage run \

README.md

@@ -65,7 +65,7 @@ provides built-in composable modules that are packaged into a few key components
 <!-- markdownlint-disable -->
 Component | Description |
 ---- | --- |
-[**physicsnemo.models**](https://docs.nvidia.com/deeplearning/physicsnemo/physicsnemo-core/api/physicsnemo.models.html) | A collection of optimized, customizable, and easy-to-use families of model architectures such as Neural Operators, Graph Neural Networks, Diffusion models, Transformer models, and many more|
+[**physicsnemo.models**](https://docs.nvidia.com/physicsnemo/latest/user-guide/model_architectures.html) ( [More Details](https://docs.nvidia.com/physicsnemo/latest/physicsnemo/api_models.html)) | A collection of optimized, customizable, and easy-to-use families of model architectures such as Neural Operators, Graph Neural Networks, Diffusion models, Transformer models, and many more|
 [**physicsnemo.datapipes**](https://docs.nvidia.com/deeplearning/physicsnemo/physicsnemo-core/api/physicsnemo.datapipes.html) | Optimized and scalable built-in data pipelines fine-tuned to handle engineering and scientific data structures like point clouds, meshes, etc.|
 [**physicsnemo.distributed**](https://docs.nvidia.com/deeplearning/physicsnemo/physicsnemo-core/api/physicsnemo.distributed.html) | A distributed computing sub-module built on top of `torch.distributed` to enable parallel training with just a few steps|
 [**physicsnemo.curator**](https://github.com/NVIDIA/physicsnemo-curator) | A sub-module to streamline and accelerate the process of data curation for engineering datasets|
@@ -119,7 +119,7 @@ PhysicsNeMo offers a library of state-of-the-art models specifically designed
 for Physics-ML applications. Users can build any model architecture by using the underlying
 PyTorch layers and combining them with curated PhysicsNeMo layers.
 
-The [Model Zoo](https://docs.nvidia.com/deeplearning/physicsnemo/physicsnemo-core/api/physicsnemo.models.html#model-zoo)
+The [Model Zoo](https://docs.nvidia.com/physicsnemo/latest/user-guide/model_architectures.html)
 includes optimized implementations of families of model architectures such as
 Neural Operators:
 
@@ -333,16 +333,14 @@ way is to start with a reference sample and then update it for your own use case
 
 ## Installation
 
-The following instructions help you install the base PhysicsNeMo modules to get
-started. In addition to this, optional dependencies can be installed to provide
-additional functionality. A complete list of optional dependencies is available
-in the [`pyproject.toml`](./pyproject.toml) file.
+You can install PhysicsNeMo in two supported ways: **via pip** (native pip or
+**uv**) or by using the **NVIDIA container image**. Choose the method that fits your
+environment and workflow.
 
-The [training recipes](./examples) are not packaged into the pip wheels or the
-container to keep the footprint low. We recommend users clone the appropriate
-training recipes and use them as a starting point. These training recipes may
-require additional example-specific dependencies, as indicated through an
-associated `requirements.txt` file in such cases.
+The following instructions cover the base PhysicsNeMo modules. Optional dependencies
+are listed in [`pyproject.toml`](./pyproject.toml). The [training recipes](./examples)
+are not bundled in the pip wheels or container; clone the repo and use the examples
+as a starting point. Many examples have a `requirements.txt` for extra dependencies.
 
 ### CUDA Backend Selection
 
@@ -363,7 +361,7 @@ via an extra that is orthogonal to the feature extras - combine them freely:
 
 ### PyPI
 
-The recommended method for installing the latest version of PhysicsNeMo is using PyPI:
+Install the latest version from PyPI:
 
 ```Bash
 pip install nvidia-physicsnemo
@@ -387,7 +385,7 @@ The installation can also be verified by running the [Hello World](#hello-world)
 
 ### uv
 
-For development or to run examples, we recommend using [uv](https://docs.astral.sh/uv/)
+For development or to run examples from source, you can use [uv](https://docs.astral.sh/uv/)
 to clone the repository and sync dependencies:
 
 ```Bash
@@ -411,7 +409,7 @@ uv sync --extra cu12 --extra nn-extras
 
 ### NVCR Container
 
-The recommended PhysicsNeMo Docker image can be pulled from the
+The PhysicsNeMo Docker image can be pulled from the
 [NVIDIA Container Registry](https://catalog.ngc.nvidia.com/orgs/nvidia/teams/physicsnemo/containers/physicsnemo)
 (refer to the NGC registry for the latest tag):
 
@@ -444,6 +442,13 @@ pip install .
 python -c "import physicsnemo; print('PhysicsNeMo version:', physicsnemo.__version__)"
 ```
 
+For editable installs, testing your changes locally, and contribution workflows,
+see the [Customizing PhysicsNeMo](https://docs.nvidia.com/physicsnemo/latest/resources/customization_guide.html)
+guide in the documentation. See also the [contributing guidelines](CONTRIBUTING.md)
+for pull requests, coding style, and CI, and the
+[developer wiki](https://github.com/NVIDIA/physicsnemo/wiki) for community and
+contributing overview.
+
 ### Building Docker from Source
 
 To build the PhysicsNeMo Docker image:
@@ -504,8 +509,12 @@ PhysicsNeMo is an open-source collaboration, and its success is rooted in commun
 contributions to further the field of Physics-ML. Thank you for contributing to the
 project so others can build on top of your contributions.
 
-For guidance on contributing to PhysicsNeMo, please refer to the
-[contributing guidelines](CONTRIBUTING.md).
+For guidance on contributing to PhysicsNeMo, see the [contributing guidelines](CONTRIBUTING.md)
+(pull requests, coding style, CI).
+For install-from-source steps, editable installs, and testing your changes, see
+the [Customizing PhysicsNeMo](https://docs.nvidia.com/physicsnemo/latest/resources/customization_guide.html)
+guide in the documentation. For community and contributing overview, see the
+[developer wiki](https://github.com/NVIDIA/physicsnemo/wiki).
 
 ## Cite PhysicsNeMo