Changelog¶
0.1.19 — 2026-04-01¶
Highlights:
- Restored automatic
laplace-torchintegration forhessian_structure="kron"and"full"when the optional dependency is installed. - Added a
laplaceextra (uqdeepnn[laplace]) so PyPI users can reproduce the older full-Hessian and Kronecker-backed Laplace behavior. - Clarified the PyPI install instructions for legacy Laplace tutorials that rely on the
laplace-torchbackend.
0.1.18 — 2026-03-28¶
Highlights:
- Added
GraphNeuralOperator2Das a grid-as-graph message-passing neural operator with optional last-layer Laplace compatibility. - Added local The Well Gray-Scott loading utilities plus a new graph-neural-operator ensemble tutorial under
notebooks/graphs/. - Extended the docs site and model architecture inventory to cover graph operators and their UQ workflows.
0.1.17 — 2026-03-12¶
Highlights:
- Expanded the VI family with heteroscedastic regression, multi-output regression, and last-layer VI wrappers.
- Added a dedicated
notebooks/vi/notebook family with five executable VI tutorials covering regression, aleatoric noise, multi-output prediction, and classification. - Strengthened the VI API docs with workflow conventions, tensor-shape guidance, and
UQResultfield mapping. - Added notebook smoke validation support for end-to-end execution checks.
0.1.16 — 2026-03-10¶
Highlights:
- Expanded the deep-ensemble method guide to document all five ensemble variants with equations, uncertainty decompositions, and paper citations.
- Added five scientific deep-ensemble tutorials: ADR regression, heteroscedastic ADR regression, elasticity failure classification, elastic-bar multi-output regression, and heteroscedastic transport regression.
- Extended ensemble tests and the package API to cover regression, classification, multi-output, and heteroscedastic ensemble variants.
0.1.15 — 2026-03-10¶
Highlights:
- Added
ConditionalUNet2DandSinusoidalTimeEmbeddingindeepuq.models.diffusionfor conditional diffusion tutorials on scientific fields. - Added
notebooks/sciml/generative/ConditionalDiffusion_Heat2D_Tutorial.ipynbfor sparse-sensor 2D heat-field reconstruction with sample-based uncertainty. - Added diffusion-model docs coverage via
docs/tutorials/sciml-conditional-diffusion-heat2d.mdanddocs/api/models/diffusion.md. - Updated the model-architecture inventory and tutorial index to include generative / diffusion models.
- Added diffusion-model unit tests and fixed the notebook diffusion schedule to avoid CPU/CUDA device mismatch during execution.
0.1.14 — 2026-03-10¶
Highlights:
- Added reusable 2D Fourier Neural Operator components:
SpectralConv2D,FNOBlock2D, andFNO2D. - Added
notebooks/sciml/operators/FNO2D_Darcy_Laplace_Tutorial.ipynbcovering a 2D Darcy-flow operator surrogate with three input fields and last-layer Laplace uncertainty maps. - Added tests for 2D FNO forward behavior and last-layer Laplace compatibility.
- Added tutorial docs and architecture-table coverage for the new Darcy operator-learning example.
- Added GIF export in the Darcy notebook to compare input fields, true pressure, predictive mean, error, and epistemic uncertainty across samples.
0.1.13 — 2026-03-10¶
Highlights:
- Refined the SciML PINN Poisson tutorial with richer interior anchors, an
Adam -> LBFGStraining schedule, and a higher-prior last-layer Laplace fit. - Updated the PINN notebook to compare
diagandblock_diaglast-layer Laplace backends and keep the more stable posterior view. - Clarified the deep-ensemble Poisson tutorial to distinguish parameter-space OOD uncertainty from physical-space data sparsity.
- Refreshed the affected notebooks and tutorial pages so the GitHub repo and docs site stay aligned with the improved notebook behavior.
0.1.12 — 2026-03-08¶
Highlights:
- Switched PyPI long-description rendering to a dedicated
README_PYPI.md. - Removed GitHub workflow badges and repo-specific overview content from the PyPI package description.
- Aligned naming in the package presentation around
Deep-UQ(docs/project),uqdeepnn(PyPI), anddeepuq(Python import).
0.1.11 — 2026-03-08¶
Highlights:
- Added
DeepEnsembleWrapperas a regression-first multi-model UQ baseline. - Added new predictive backbones:
CNNRegressor2D,ResNetRegressor2D,UNet2D,UNet3D,PINN1D, andPINN2D. - Added the new model architecture inventory page plus API docs for ensembles, spatial models, and PINNs.
- Added new SciML tutorial notebooks for deep ensembles, CNN / ResNet heat surrogates, U-Net diffusion surrogates, and PINN Poisson problems.
0.1.10 — 2026-03-08¶
Highlights:
- Added reusable 3D Fourier Neural Operator components:
SpectralConv3D,FNOBlock3D, andFNO3D. - Added
notebooks/sciml/operators/FNO3D_Heat_Laplace_Tutorial.ipynbcovering a 3D periodic heat-equation surrogate with slice-based uncertainty visualization. - Added tests for 3D FNO forward behavior and last-layer Laplace compatibility.
- Added tutorial docs and site navigation for the new 3D SciML notebook.
0.1.9 — 2026-03-08¶
Highlights:
- Added
DeepONet1Dfor fixed-grid 1D operator-learning workflows that remain compatible withLaplaceWrapper. - Added
notebooks/sciml/operators/DeepONet_Poisson1D_Laplace_Tutorial.ipynbcovering a 1D Poisson problem with sparse forcing sensors, residual DeepONet training, and last-layer Laplace uncertainty bands. - Extended DeepONet tests to cover the 1D model and last-layer Laplace compatibility.
- Expanded the 2D Burgers SciML notebook explanations and comments for easier first-time reading.
- Added tutorial docs and site navigation for the new 1D SciML notebook.
0.1.8 — 2026-03-07¶
Highlights:
- Added CI workflows for tests (
tests.yml) and quality checks (lint.yml). - Added tag-driven PyPI release workflow (
release.yml) with trusted publishing. - Introduced standardized uncertainty container
deepuq.UQResult. - Added non-breaking
predict_uqAPIs across methods and GP models. - Added manual multi-dataset benchmark suite under
benchmarks/. - Added tracked-data policy docs and large-file guard script.
- Added packaging extras in
pyproject.toml(dev,tests,docs,benchmarks,notebooks). - Updated docs links and usage/examples for the unified uncertainty API.
- Added
DeepONet2Dfor operator-learning experiments in scientific machine learning. - Added a new
notebooks/sciml/operators/DeepONet_Burgers_Laplace_Tutorial.ipynbtutorial covering 2D viscous Burgers operator learning with Laplace uncertainty. - Added tests for DeepONet forward behavior and Laplace last-layer compatibility.
- Added tutorial docs and navigation for the new SciML notebook section.
0.1.4 — 2026-03-03¶
Highlights:
- Fixed homepage rendering behavior for GitHub Pages deployment.
- Added cache-busted docs assets (
extra-v2.css,extra-v2.js) to avoid stale browser rendering. - Removed stale
laplace-torchlanguage from Laplace notebooks/examples.
0.1.3 — 2026-03-03¶
Highlights:
- Removed external
laplace-torchdependency from package requirements. - Added native
kronandfullLaplace backends underLaplaceWrapper. - Updated Laplace docs/usage notes to reflect native support for all Hessian structures.
- Refreshed Laplace Hessian comparison tutorial updates.
0.1.2 — 2026-03-03¶
Highlights:
- Expanded Laplace support through
LaplaceWrapper: - native
diag,fisher_diag,lowrank_diag,block_diag - native
kronandfullimplementations (nolaplace-torchdependency) - Added Laplace notebooks under
notebooks/laplace/: - full-Hessian tutorial
- multi-structure comparison tutorial
- Improved VI/Laplace tutorial consistency and diagnostics
- Refined README method table and package documentation links
0.1.1¶
- Packaging and release updates for PyPI publication
0.1.0¶
- Initial public package release