.. _releasenotes-70: """""""""""""""""""""""""""" Version 7.0 (2026/11/17) """""""""""""""""""""""""""" **Highlight: Python API** Dakota 7.0 introduces a Python API that allows studies to be constructed and run programmatically without writing a Dakota input file. Study components — methods, models, variables, interfaces — are ordinary Python or C++ objects, enabling Dakota to be embedded inside larger applications, automation pipelines, and agentic AI workflows. Python bindings are provided for nearly all iterators and models, including parameter studies, UQ methods, calibration, optimization, surrogate-based methods, ensemble surrogates, and meta-iterators. The APIs share the same configuration grammar as JSON input but replace pointer-based ``id_*`` / ``*_pointer`` relationships with direct object references. Kwargs and Pydantic model configurations are both supported for setting keyword options. The :ref:`dakota-to-json ` command-line tool and its backing Python module ``dakota.tools.freeform_to_json`` can convert an existing freeform Dakota input to a JSON file that Dakota accepts directly with ``dakota -json`` and that can serve as a starting point for a Python-based study. To simplify the API, a few noteworthy changes were made to Dakota's input spec. These changes affect not only the API and JSON input formats, but also the freeform format. First, the four surrogate types (global, local, multipoint, ensemble) were promoted to model types. Instead of ``model surrogate global``, it's now necessary to specify ``model global_surrogate``. Similarly, ``method hybrid sequential`` was promoted to ``method hybrid_sequential``. *Enabling / Accessing:* * Follow the instructions to :ref:`install a pre-built package `, including setting your ``PYTHONPATH``. Dakota's compiled Python modules were built using Python 3.13, and compatible only with this Major.minor version. * Or, build Dakota with the ``DAKOTA_PYTHON_STUDY`` CMake option enabled, and build the ``install-python`` target to install Dakota's Python package into your environment. A minimum of Python 3.10 is required. *Documentation:* :ref:`library-api` in the User's Manual (construction guide, tutorial, and Python reference). **Highlight: Rewritten Introduction and Tutorial** The introductory and tutorial sections of the User's Manual have been completely rewritten. The new material covers how to run your first Dakota study step by step, how to connect your own simulation code as an analysis driver, and how the main study types differ, all with examples in both JSON and freeform input formats. Existing users will find the restructured content easier to navigate; new users get a more direct path from installation to a working study. *Documentation:* :ref:`helloworld-main` and :ref:`couplingtosimulations-main` in the User's Manual. **Improvements by Category** *Interfaces, Input/Output* - JSON input files are now checked for required fields and array item types, with improved validation of union member constants and alternatives. - Integer fields in both JSON and free-form input now require whole-number values. - The JSON parser now correctly handles ``true``/``false`` and ``{}`` based on the declared schema type. - ``import_points`` results are now correctly archived in HDF5 output. - Dakota will now raise an error when two concurrent studies would produce colliding results output filenames. - ``ResultsManager`` and ``EvaluationStore`` instances are now study-specific, enabling safe concurrent use in library API contexts. - Fixed an HDF5 close issue that could leave output files in an inconsistent state. *Models* - Surrogate model types are now promoted to first-class model types in the input grammar and documentation (e.g. ``model surrogate global`` is now ``model global_surrogate``. - Fixed: honor per-component ASV for field approximations. - Fixed: out-of-range model form in ``NonDExpansion`` online cost averaging. *Optimization Methods* - **ROL** (Rapid Optimization Library): Updated to use Trilinos 16.2. The old monolithic ROL interface has been removed; the modular interface is now the only supported path. Evaluation cache management moved to ``ROLOptimizer``; several memory errors identified by valgrind have been fixed. - **EGO** (Efficient Global Optimizer): Fixed a bug where merit-function updates after convergence corrupted best-point selection when nonlinear constraints are present (contributed by Werner Van Geit). - Fixed: CONMIN constraint maps are now reset correctly between multiple successive runs. - Fixed: NOMAD best-point loop now correctly indexes ``discRealVars[j]``. - Fixed: In least-squares calibration with nonlinear constraints and scaling enabled, ``LeastSq::post_run`` was unscaling best-point constraints from ``iter_resp`` (the model's last evaluation) rather than from the solver-cached best response in ``best_resp``, overwriting correct best constraints with stale values. The fix stages the scaled best constraints from ``best_resp`` into ``iter_resp`` before the in-place unscale. *UQ Methods* - **Adaptive Sampling**: Fixed multiple pre-existing correctness bugs including constant-liar pop count, duplicate batch picks, and topology fallback distances. Asynchronous evaluation support added. - Fixed: In the per-response-file branch of ``ExperimentData::load_experiment``, omitting ``experiment_variance_type`` left ``sigma_scalars`` zero-filled, caching a zero covariance determinant (log det = −∞) that corrupts the Bayesian log-likelihood for every sample. Per-response-file inputs now default to unit variance, matching the scalar-data-file branch. - Fixed: ``Variables::load`` now validates each array length from the restart archive against the ``SharedVariablesData``-derived counts and aborts with an error on mismatch, preventing out-of-bounds Teuchos views from a restart file whose variable counts are inconsistent with the current input. - Fixed: In ``NestedModel::iterator_response_overlay``, the identity-map branch for secondary (constraint) gradients was writing ``mapped_grad[m_index]`` inside the loop over derivative variables ``j``, leaving most gradient entries zero and writing out of bounds when ``m_index >= num_mapped_deriv_vars``. The index is now ``mapped_grad[j]``, matching the primary identity branch. - Fixed: In ``Constraints::reshape_update_linear``, contracting from an ALL view to a non-ALL view with both linear inequality and equality constraints present used two separate ``cv``-section loops that shared ``curr_cntr``/ ``prev_cntr``, causing the equality matrix to be copied from the wrong columns and indexing past ``num_av``. Both matrices are now copied in a single loop per column section. *Solution Verification* - Fixed: ``RichExtrapVerification::converge_order`` and ``converge_qoi`` initialized the convergence tolerance, iteration counter, and (for ``converge_order``) ``prev_order`` once outside the per-factor loop. After the first factor converged, all subsequent factors skipped refinement entirely and ``extrapolate_result`` reported uninitialized ``convOrder``/``extrapQOI``/``numErrorQOI`` values. Each factor now resets these variables independently at the start of its refinement loop. *Sensitivity Analysis* - The :dakkw:`method-sampling-variance_based_decomp-vbd_sampling_method-binned` approach to variance-based decomposition now works with discrete variables. *Calibration* - Documentation added explaining how Dakota computes confidence intervals on calibration parameters, including derivation, numerical computation, and assumptions. See the NLS confidence intervals section of the Theory Manual. *Hybrid Methods* - ``hybrid_collaborative`` and ``hybrid_embedded`` method types have been removed. Hybrid sub-methods are now promoted to first-class method types and should be specified directly. Use ``hybrid_sequential`` (keyword ``method > hybrid > sequential``) for sequential hybrid strategies. **Miscellaneous Enhancements and Bugfixes** - DOT optimizer: bumped to version 3.71; ALPSAV state is now saved between calls, correcting behavior in constrained problems. - Fixed uninitialized pointer issues identified across several model constructors. - Fixed ``pyprepro`` function name typo. - ``ext_python`` module moved from ``src/`` to the ``python/`` package. - Resolved duplicate-RPATH install errors and build-tree RPATH issues on macOS. - Fixed Python API shared library loading on macOS and Linux. - ``libpython`` is no longer linked into ``libdakota_surrogates`` on macOS (resolves JNI linking issues). - ``EXT_TPL_LIBS`` is now linked into ``libdakota_src.so`` to enable Python import. - Added a ``make install-python`` CMake target that installs the Dakota Python package into the active Python environment (run ``make install`` first). - ``dakota_parser`` now inherits C++ compiler flags from the parent build and probes for Python 3 automatically. - Legacy input ``Data`` classes have been removed; construction throughout Dakota now uses the ``IRStore`` path, laying the foundation for fully decoupled library-API component construction. - ``DL_Solver`` dynamically loadable solver interface has been removed. **Deprecated and Changed** - **NIDR parser removed.** The ``-parser legacy`` command-line argument is no longer available. All input, whether free-form or JSON, is processed by the new parser. - **``hybrid_collaborative`` and ``hybrid_embedded`` removed.** Workflows using these types must be migrated to use their constituent methods directly, or to ``hybrid_sequential``. - **C++20 required.** The minimum C++ standard for building Dakota has been raised from C++17 to C++20. - **Boost minimum raised to 1.75.** - **Trilinos updated to 16.2** (required when ROL is enabled). **Compatibility** - A C++20-capable compiler is now required to build Dakota from source. - Boost 1.75 or newer is required. - Trilinos 16.2 or newer is required when building with ROL enabled. - Python 3.9 or newer is needed to build Dakota. - Python 3.10 or newer and Pydantic 2.12 or newer are required to regenerate or modify Dakota's input grammar (``python/dakota/spec/``). - The ``-parser legacy`` command-line flag has been removed; input files that relied on NIDR-specific parsing behavior may need minor adjustment. **Other Notes and Known Issues** - The library APIs are public and supported but are still evolving; interfaces may change in upcoming releases. - Support for plugging in user-supplied methods, models, and surrogates from third-party libraries is planned for a future release, as is availability of the Python API through PyPI.