.. _releasenotes-622:

"""""""""""""""""""""""""
Version 6.22 (2025/05/15)
"""""""""""""""""""""""""

**Improvements by Category**

*Interfaces, Input/Output*

- The :ref:`pyprepro template processing <interfaces:dprepro-and-pyprepro>` tool now supports
  :ref:`JSON format Dakota parameters files <variables:parameters:json>`.

*Optimization Methods*

- :dakkw:`ROL optimizer <method-rol>` was updated to release 2.0.

*UQ Methods*

- Multifidelity sampling methods now summarize the final variance metrics
  per quantity of interest (QoI) and report 95% confidence intervals for
  each of the QoI means.

- Multifidelity sampling methods now admit different reductions for forming
  a scalar optimization metric in multiple QoI cases: max, average, and
  p-norm can now be used during sample allocation solves.  Refer to :dakkw:`optimization metric <method-approximate_control_variate-solver_metric>`.
 
- ML BLUE now employs analytic gradients of the estimator variance to improve
  solution robustness.

- All multifidelity sampling methods (previously only ML BLUE, but now
  including all ACV and generalized ACV solves) now employ the same
  truncated SVD mitigations for ill-conditioned matrix solves.
  Previous Cholesky factorization approaches have been retired.

- Support for MUQ's :dakkw:`multilevel MCMC algorithm <method-bayes_calibration-muq-multilevel_mcmc>` was added.


**Miscellaneous Enhancements and Bugfixes**

- Enh: Dakota now uses GoogleTest instead of Boost.Test
- Enh: Dakota now uses std::filesystem when available instead of Boost.filesystem.
- Enh: The Trilinos snapshot was updated to version 16.1
- Enh: Switch offline_pilot modes to use oracle covariances when computing control variate parameters, for better consistency in final roll-up between model-based and group-based approaches.
- Enh: Refactor multilevel_multifidelity_sampling to modernize and improve code reuse
- Enh: Weighted MLMC sampling now accepts specification of an 
  :dakkw:`optimization strategy <method-multilevel_sampling-weighted>`.
- Bug fix: In generalized ACV, deduct sunk pilot cost for inactive models from budget when solving for sample allocations for an active subset.  In ML BLUE, a similar fix is applied for either inactive models (shared pilot) or inactive groups (independent pilot).
- Bug fix: In generalized ACV, reset the optimal merit function for each search over model subsets and DAGs (important for online_pilot iteration and model tuning)
- Bug fix: Hardening of model selection within multilevel_sampling and multifidelity_sampling (hierarchical cases promoted to generalized ACV)
- Bug fix: Small fixes to JEGA to satisfy recent versions of clang (issues `#178 <https://github.com/snl-dakota/dakota/issues/178>`_ and `#94 <https://github.com/snl-dakota/dakota/issues/94>`_)
- Bug fix: Boost version check fixed (issue `#163 <https://github.com/snl-dakota/dakota/issues/163>`_)
- Bug fix: Broken links in examples (issue `#162 <https://github.com/snl-dakota/dakota/issues/162>`_)

**Deprecated and Changed**

- The legacy Python interface, which was incompatible with NumPy 2, has been removed.

**Compatibility**

- Dakota now requires C++17 and CMake 3.23
- Legacy Python 2 support has been removed, and Python 3 is required for all optional python dependencies