Skip to content

Changelog

0.1.19 — 2026-04-01

Highlights:

  • Restored automatic laplace-torch integration for hessian_structure="kron" and "full" when the optional dependency is installed.
  • Added a laplace extra (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-torch backend.

0.1.18 — 2026-03-28

Highlights:

  • Added GraphNeuralOperator2D as 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 UQResult field 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 ConditionalUNet2D and SinusoidalTimeEmbedding in deepuq.models.diffusion for conditional diffusion tutorials on scientific fields.
  • Added notebooks/sciml/generative/ConditionalDiffusion_Heat2D_Tutorial.ipynb for sparse-sensor 2D heat-field reconstruction with sample-based uncertainty.
  • Added diffusion-model docs coverage via docs/tutorials/sciml-conditional-diffusion-heat2d.md and docs/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, and FNO2D.
  • Added notebooks/sciml/operators/FNO2D_Darcy_Laplace_Tutorial.ipynb covering 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 -> LBFGS training schedule, and a higher-prior last-layer Laplace fit.
  • Updated the PINN notebook to compare diag and block_diag last-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), and deepuq (Python import).

0.1.11 — 2026-03-08

Highlights:

  • Added DeepEnsembleWrapper as a regression-first multi-model UQ baseline.
  • Added new predictive backbones: CNNRegressor2D, ResNetRegressor2D, UNet2D, UNet3D, PINN1D, and PINN2D.
  • 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, and FNO3D.
  • Added notebooks/sciml/operators/FNO3D_Heat_Laplace_Tutorial.ipynb covering 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 DeepONet1D for fixed-grid 1D operator-learning workflows that remain compatible with LaplaceWrapper.
  • Added notebooks/sciml/operators/DeepONet_Poisson1D_Laplace_Tutorial.ipynb covering 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_uq APIs 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 DeepONet2D for operator-learning experiments in scientific machine learning.
  • Added a new notebooks/sciml/operators/DeepONet_Burgers_Laplace_Tutorial.ipynb tutorial 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-torch language from Laplace notebooks/examples.

0.1.3 — 2026-03-03

Highlights:

  • Removed external laplace-torch dependency from package requirements.
  • Added native kron and full Laplace backends under LaplaceWrapper.
  • 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 kron and full implementations (no laplace-torch dependency)
  • 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