scirex 1.0.0__tar.gz

scirex-1.0.0/.gitignore

@@ -0,0 +1,138 @@
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+*.egg
+*.egg-info/
+dist/
+build/
+eggs/
+parts/
+var/
+sdist/
+wheels/
+*.egg-link
+.installed.cfg
+lib/
+lib64/
+
+# Virtual environments
+.env
+.venv
+env/
+venv/
+ENV/
+
+# Distribution / packaging
+.Python
+develop-eggs/
+downloads/
+.eggs/
+MANIFEST
+
+# Testing
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+htmlcov/
+
+# Jupyter
+.ipynb_checkpoints
+*.ipynb
+
+# JAX / XLA
+xla_cache/
+.xla_cache/
+*.ptxas
+jax_cache/
+jax_profiles_epoch/
+
+# Runtime logs
+neighbor_logs/
+
+# Checkpoints / model artifacts
+checkpoints/
+*.pt
+*.pth
+*.ckpt
+*.orbax/
+wandb/
+
+# Data
+*.h5
+*.hdf5
+*.nc
+*.zarr/
+*.tgz
+/data/
+# Downloaded benchmark datasets (default data_root for the WNO Zenodo loaders)
+/scirex/
+
+# AirfRANS dataset (raw + cached .npz features) and its Transolver run
+# outputs -- /Output* (not just /Output/) covers the many ad-hoc per-run
+# variants (--output-dir overrides) these scripts produce; Dataset_cache/
+# and scripts/logs/ are already ignored above.
+/Dataset/
+/Output*/
+
+# IDEs
+.vscode/
+.idea/
+*.swp
+*.swo
+.DS_Store
+
+# Docs build
+site/
+doc/_build/
+
+# uv
+.uv/
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# ruff
+.ruff_cache/
+
+
+# Benchmark model checkpoints (regenerated by training runners)
+scripts/benchmarks/checkpoints/
+experiments/
+
+# Li et al. Darcy .mat files (large, fetched manually) read by
+# scripts/benchmarks/darcy/
+src/scirex/data/darcy_wno_paper_cache/
+
+# VTK output files (large; regenerated by the benchmarks' infer.py)
+scripts/benchmarks/vtk_outputs/
+
+# Status / project management documents (local only)
+status/
+
+# Local MLflow / Optuna stores (scirex.tuning with a sqlite or file backend)
+mlruns/
+mlartifacts/
+mlflow.db
+optuna.db
+
+Dataset_cache/
+scripts/logs/
+.jax_cache/
+
+# training logs
+logs/
+
+# Output of scripts/reproduce.py runs (contains local paths and logs)
+reproduce_out/

scirex-1.0.0/CHANGELOG.md

@@ -0,0 +1,68 @@
+<!--
+SPDX-FileCopyrightText: 2026 ZenteiQ AiTech Innovations Pvt. Ltd.
+SPDX-License-Identifier: CC-BY-4.0
+-->
+
+# Changelog
+
+All notable changes to **ScireX** are documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [1.0.0] - 2026-11-12
+
+First release of ScireX: a JAX/Flax library of neural operators for scientific machine
+learning, succeeding the original [SciREX](https://github.com/zenteiq-sciml/SciREX) research
+framework (a separate project from the AiREX Lab at IISc). It has a defined public API: the
+stable names, the three stability tiers, the versioning rules and the deprecation policy are
+in [`doc/dev_guide/api_policy.md`](doc/dev_guide/api_policy.md), and the public names are
+listed in the [API reference](doc/api/index.md).
+
+### Highlights
+
+- **Models:** FNO, WNO, GINO and Transolver in pure JAX/Flax, plus the experimental GIWO.
+- **Physics-informed training:** the finite-element Ritz-energy loss (`Q4Galerkin`), which
+  trains a model without labelled solutions. Equation-residual losses are experimental
+  (`scirex.experimental`).
+- **Benchmarks:** Darcy flow (data-driven and physics-only), pipe flow, 2D Navier-Stokes,
+  lid-driven cavity, chip cooling, AirfRANS, ShapeNetCar and Car-CFD, at a matched parameter
+  budget, with write-ups, dataset guides and the commands to reproduce each result.
+- **Checkpoints:** the trained benchmark checkpoints are on Hugging Face, one repository per
+  checkpoint (`zenteiq-ai/scirex-<model>-<problem>`). Lid-driven cavity has none because its
+  data is not public.
+- **Your own data:** `scripts/train.py` and `scripts/predict.py` train and run any of the four
+  models on your own data in one command (see the *Use It* page).
+- **Install and check:** pip (CPU, CUDA 12, TPU), Docker and source installs, with
+  `python -m scirex.info` and `scripts/dev/check_platform.sh` to check a machine.
+- **Documentation and tooling:** an MkDocs site with the API reference, theory guides and
+  developer guides; hyperparameter tuning with Optuna and MLflow (`scirex.experimental.tuning`);
+  Flax-native checkpointing and Weights & Biases logging; CI on Python 3.10, 3.11 and 3.12.
+
+### Deprecated
+
+These names still work and emit a `DeprecationWarning`. They are removed in 1.2.0 at the earliest.
+
+| Name | Use instead |
+|---|---|
+| `scirex.training.append_grid_nhwc`, `append_grid_nwc`, `create_train_state`, `make_schedule`, `per_frame_relative_l2`, `plot_loss_curves`, `relative_l2_sum`, `resolve_dtype`, `write_metrics` | Not public API. They live in `scirex._internal.wno_train_utils` |
+| `scirex.losses.ADDiff`, `AdvectionTerm`, `AdvectionTermAD`, `BatchedVariableCoefficientGalerkinResidualLossT3`, `BatchedVariableCoefficientGalerkinResidualLossT6`, `DarcyHybridLossGINO`, `DiffusionTermAD`, `DivergenceTerm`, `FourierDiff`, `GradientTerm`, `NavierStokesEqnLoss`, `PDEResidualLoss`, `ReactionTerm`, `ReactionTermAD`, `SourceTermAD`, `T3Galerkin`, `T6Galerkin`, `TransientTerm` | Experimental. Import them from `scirex.experimental.losses` |
+| `scirex.training.SimpleDataLoader`, `Trainer`, `load_training_state`, `save_training_state` | Experimental. Import them from `scirex.experimental.training` |
+| Module paths `scirex.data.web_utils`, `scirex.layers.common.einsum_utils`, `scirex.layers.common.segment_csr`, `scirex.training.torch_setup`, `scirex.training.trainer`, `scirex.training.training_state`, `scirex.training.wno_train_utils`, `scirex.distributed`, `scirex.geometry`, `scirex.training.patching`, `scirex.tuning`, `scirex.config.distributed`, `scirex.models.giwo` | The modules moved to `scirex._internal.*` (private) and `scirex.experimental.*`. The old path still imports the same module |
+
+### Known limitations
+
+- The tensor-GaLore projector (`scirex.training.optim`, not part of the stable API) cannot use
+  `jax.checkpoint` or a warm restart with Tucker factors, because they call non-JAX TensorLy
+  functions. Three tests mark this as an expected failure (`xfail`, six runs on CPU and GPU).
+- `MeshDataModule` normalisation divides by `max - min`; a feature that is constant over the data
+  gives a `RuntimeWarning` and NaN values.
+- VTK 9.6 deprecates `GetData` on cell arrays. The AirfRANS loaders still use it and emit a
+  `DeprecationWarning`. The `benchmark` extra allows `vtk>=9.0` with no upper bound, so a later
+  VTK release may remove the method.
+- Plain `optax` optimizers (e.g. `optax.adamw`) do not raise on complex-valued parameters; they
+  run without error but update the imaginary parts in the wrong direction (see
+  `scirex.training.optim.adamw_cx`, which handles this correctly). Nothing currently guards
+  against building a plain-Optax optimizer over complex parameters by mistake.
+
+[1.0.0]: https://github.com/zenteiq-research/scirex/releases/tag/v1.0.0

scirex-1.0.0/CITATION.cff

@@ -0,0 +1,20 @@
+cff-version: 1.2.0
+message: "If you use ScireX in research, please cite the software and the relevant method papers."
+title: "ScireX"
+abstract: "Production-grade JAX/Flax implementations of Neural Operators for scientific machine learning."
+type: software
+authors:
+  - name: "ZenteiQ AiTech Innovations Pvt. Ltd."
+    email: research@zenteiq.ai
+version: "1.0.0"
+license: "LGPL-3.0-only"
+repository-code: "https://github.com/zenteiq-research/scirex"
+url: "https://github.com/zenteiq-research/scirex"
+keywords:
+  - Neural Operators
+  - Scientific Machine Learning
+  - Physics Informed Operator Learning
+  - JAX
+  - Flax
+  - PDEs
+date-released: "2026-11-12"