Skip to content

CHANGELOG.md

Latest commit

 

History

History
151 lines (110 loc) · 17.6 KB

CHANGELOG.md

File metadata and controls

151 lines (110 loc) · 17.6 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

1.1.6 - 2026-08-26

Added

  • Input validation now rejects an odd population size (n_mdl) when the GA tournament selection style (ga_type='tournament') is chosen, since tournament pairing silently drops a model on odd populations.
  • The MESS input reader (readers/mess_input.py) now rejects any Well, Bimolecular, Barrier, or Fragment name containing the reserved parameter-name separator __ (dbs), logging an error and stopping the run gracefully. This protects the <item>__<ptype> naming contract that all parameter-type parsing relies on.

Fixed

  • Parameter-type (Ptype) identification strengthened throughout the code.
  • Genetic-algorithm convergence of multiplicative parameters is now evaluated with geometric means and std.

1.1.5 - 2026-08-25

Fixed

  • Multiplicative-parameter perturbation, trust boundaries, and derivative steps are now computed in log space, fully consistent with the (already log-space) theory scoring. In the multiplicative branch only (if, sfc, mrc, bfc, frequencies), the arithmetic factor 1 + (std - 1) * max_std / 1 + (uc - 1) * step is replaced by the geometric/power factor std**max_std / uc**step: get_boundaries returns (value / std**max_std, value * std**max_std), get_scale uses log-space sigma ln(uncertainty), and the SA/Nelder-Mead calculate_dstep uses factor uc**step (SA step = lin_fact / sensi_d, NM step = nm_dstep). As a result the theory score at the trust boundary equals max_std**2 exactly, and the lognormal log-space sampling sigma ln(uncertainty) yields ±2/3/4 sigma coverage of 95.45/99.73/99.99%. Additive and percentage parameters and scoring.py are unchanged.

Changed

  • The multiplicative trust region is now wider than before for uncertainty > 1 (intended correction of the above reconciliation). Prior multiplicative-run results are therefore not bit-for-bit comparable.

1.1.4 - 2026-08-25

Added

  • New user-settable input keyword q_name (default day-long-cpu) exposing the queuing system's queue/partition as a free-form string passed to SLURM via #SBATCH -p. Registered in default_settings, read by QueueingSystem, and surfaced in the launcher GUI (Resources section, with save/load round-trip) and the dashboard metadata.

Fixed

  • The direction-dependent sensitivity-analysis and Nelder-Mead derivative step now handles multiplicative parameters (if, sfc, mrc, bfc, frequencies) with a truly multiplicative, direction-dependent step: with factor f = 1 + (uc - 1) * sensi_d (Nelder-Mead uses nm_dstep), the up step (side +1) is value * f and the down step (side -1) is value / f (e.g. uc=1.1, sensi_d=0.1 → steps [1/1.01, 1.01]). Previously the step keyed off the parameter's log-normal distribution rather than its class and applied a log-space scale additively, producing a wrong / symmetric step for multiplicative parameters. Additive and percentage parameters (value + scale * sensi_d * side) are unchanged. This is the perturbation-step counterpart of the log-space multiplicative scoring fix in 1.1.3.
  • Genetic-algorithm convergence (actualize_conv) now measures multiplicative-parameter convergence in log space (|ln(old / new)| for both the mean and the standard deviation) against param_conv, instead of a percent-style relative change. Percentage and additive parameters are unchanged.

1.1.3 - 2026-08-24

Fixed

  • Per-parameter specific_std overrides now govern the perturbation boundaries (the trusted range), not only the sampling scale and scoring weight.
  • Theory-score contribution for multiplicative parameters (if, sfc, mrc, bfc, frequencies) is now computed in log space as (ln(value / reference) / ln(uncertainty))**2, replacing the previous linear distance/scale. The penalty is now symmetric under a factor f versus its inverse 1/f and consistent with the perturbator's log-normal (log-space) sampling of those parameters. Additive and percentage parameters are unchanged. This is the scoring-side counterpart of the log-space perturbation correction shipped in 1.1.2.

1.1.2 - 2026-08-24

Fixed

  • Fixed a bug where the postprocessing was reading rates from files on disk even when the rates are in db, potentially ready from MESS output calculated on a different P/T grid. This was causing a crash with an error P not in list, with P being the value in the file on disk not being in the postprocess conditions list. This is now bypassed, and the rates already in DB are always read from the DB.
  • CI now installs the agentic extra (pip install -e .[test,agentic]) so the agentic-pipeline tests (test_agentic_pipeline_ci.py) are collected and run, instead of aborting collection with ModuleNotFoundError: No module named 'anthropic'.
  • Multiplicative-parameter log-normal perturbation and the asymmetric sensitivity-analysis / Nelder-Mead derivative steps were moved off a value-dependent, additive log-space treatment onto a value-independent log-space sigma so that ±max_std·σ coincides with the perturbation boundaries; removed dead get_mean_sigma. (Superseded in Unreleased: the multiplicative sigma is now ln(uncertainty) and the boundary factor is uncertainty**max_std rather than the earlier log(1 + (std - 1) * max_std) / max_std sigma and 1 + (std - 1) * max_std boundary.)

Changed

  • Error log entries in the KiMecO logfile now include the full Python traceback. Every backend try/except that logged its error through KMOLogger now passes exc_info=True, so the traceback is appended after the message (log-line format unchanged). Postprocessing GOAT-load failures and well.py uncertainty-parsing errors now log with a traceback instead of writing to stderr / a bare print.
  • Perturbation distribution validation is now enforced per parameter category. Multiplicative parameters (if, sfc, mrc, bfc, and individual/batch frequencies) accept only log-normal or log-uniform, while additive (we, be, pow) and percentage (hrs, sigma, epsilon, fact) parameters accept only uniform or normal. The backend now hard-fails invalid category/distribution combinations (previously only the additive class was checked), and the GUI perturbation dropdowns present only the valid distributions for each category.

1.1.1 - 2026-08-04

Added

  • In the example folder, the Analysis notebook now also shows how plot the extrapolated results (rate coefficients and concentration profiles).
  • Agentic delivery pipeline for repository development: a multi-stage subagent workflow (clarification, scope assessment, planning, spec review, boundaries, CI testing, version control) coordinated by a workflow orchestrator. It ships both as Claude Code subagents under .claude/agents/ (with .claude/settings.json) and as a standalone Python/Claude-API implementation under agentic_pipeline/, runnable from the repository root via python -m agentic_pipeline.cli "<request>".
  • New optional dependency group agentic in pyproject.toml (anthropic>=0.69, pydantic>=2, pyyaml) providing the packages needed to run the Python agentic pipeline (pip install -e .[agentic]).
  • New public query helper SIM_DB.get_exp_for_table(exp_id, table) returning, for a given experiment id and generation table, a list of (profile.T, species) for every model in that table (read-only accessor for analysis/plotting of postprocessed/extrapolated experiment profiles).
  • New public accessor GOATs.get_exp_for_gen(exp_id, gen) returning, for a given experiment id and GOAT generation snapshot, a list of (profile.T, (table, mdl_id)) by resolving each ensemble member to its native generation table ({prefix}{gen:04d}) via prepare_batch_select/batch_select; this is the public API for retrieving an optimized-ensemble experiment's profiles across the members' native tables (needed for extrapolation analysis).

1.1.0 - 2026-08-04

Added

  • TimeProfile data/error CSVs now accept an optional bracketed time unit on the first-column header (e.g. time[s], TIME [ms], time[1e-3s], time[1e-3]). The time token is case-insensitive and whitespace tolerant; Cantera time units plus ms/millisecond(s) aliases and numeric-factor forms are supported, with seconds assumed when no bracket is given. A new TimeProfile.time property exposes the seconds-normalized time grid.
  • New public accessor GOATs.get_goat_param_values(gen, cols) returning dict[str, np.ndarray] of the requested SOP columns for a generation, in GOAT token order, without reconstructing models or running scoring (gen == -1 selects the last generation; out-of-range raises IndexError).
  • Each optimizer now exposes a class-level prefix attribute (GeneticAlgorithm='G', NelderMead='NM', NelderMeadSwarm='NMSG') recoverable without instantiation, and a settings→optimizer-prefix resolver drives the postprocessing table and GOATs-ensemble prefix.
  • Model and SOP objects now support value equality and hashing. Two Models are equal when they share the same SOP, status, generation and id (hash derived from the SOP parameters, generation and id); two SOPs are equal when their parameters_names are identical.

Changed

  • Postprocessing/extrapolation now writes results into the primary run databases (KMO_DB_SOP / KMO_DB_KIN / KMO_DB_SIM) instead of separate extrapolation databases. Extrapolated rate coefficients and simulations are stored in the same per-generation tables where the model was originally created ({optimizer_prefix}{gen:04d}, e.g. G0003, NM0002, NMSG0001); the GT token (GOATs ensemble) now resolves to the originating optimizer's prefix (e.g. G for the genetic algorithm), so GT and X never appear as table names.
  • Extrapolation now reuses already-computed rate coefficients: if a postprocessing experiment's (P, T) already exists in the model's KIN table, MESS is not re-run for it, and only missing (P, T) conditions are computed and appended. The postprocessing simulation is always run and saved because the initial composition differs.
  • Postprocessing simulations are appended into the existing SIM tables with banded experiment ids (offset past the original run's experiments), so the original run's simulation results are never overwritten.
  • The postprocessing log (set_postprocessing) now prints the metadata of each pp_experiment (type, temperature, pressure, species, composition) instead of the flat pp_temp / pp_pres grids.
  • The kmoui dashboard now reads extrapolated simulations from the unified SIM database; extrapolated experiments are labelled Extrapolated (band b) — <experiment metadata> in the simulations and database views.
  • The SOP GUI plotting subsection ("Type of parameter to plot" → Plot) now fetches each selected generation's data once for all selected columns instead of reconstructing full SOP/Model objects and running scoring per parameter, making parameter plotting much faster. Plotted values (including the Score parameter) are byte-identical to before and the UX is unchanged (one overlaid-histogram figure per selected column).
  • Internal GOATs.get_goat_for_gen row matching reduced from O(n²) to O(n) via an id→row map, and GOATs.get_p_for_gen optimized in place; observable behavior is unchanged (rows now returned in deterministic GOAT token order, identical shapes/dtypes and error contracts).
  • API note: database.sop_db.batch_select_cols now returns an id-keyed dict of the form {table: {row_id: (col_values...)}} (the row id is included in each entry) and no longer emits an empty .where() clause.
  • TimeProfile time grids are normalized to seconds on read (species columns untouched), and data/error files may declare different time units as long as their converted-seconds grids match (compared with a numerical tolerance).
  • In the GUI KIN section, reaction pair selection now uses only wells and bimolecular species (fragments excluded), labels entries as NAME [PES XX], enforces same-PES From/To pairing with reciprocal filtering and auto-clear of invalid selections, and blocks invalid cross-PES plotting with an explanatory message.
  • In the GUI SIM section ("Concentration profiles"), selection is now driven by a single multi-select experiment dropdown instead of separate pressure/temperature/species controls. Each entry is labelled with its experiment metadata (for TimeProfile experiments: pressure converted Pa→bar with unit, temperature in K, and the measured species, e.g. Time profile #3 — 1.013 bar, 300 K — A, B; other experiment types fall back to {exp_type} #{id}). Selecting an experiment automatically produces a separate figure per measured species, each heading naming both the species and the experiment. The change is GUI-layer only; experiment classes, the SIM-DB schema, and settings keys are unchanged.
  • A GOAT-load failure in postprocessing (set_postprocessing) now surfaces loudly: instead of silently continue-ing past a token, it prints the traceback and raises ValueError, so a failed ensemble/band is no longer silently dropped.

Removed

  • The separate postprocessing databases PP_DB_KIN.db and PP_DB_SIM.db are no longer created; extrapolation results now live in the primary run databases. The PP simulation source in the kmoui dashboard is removed accordingly.
  • The X-prefixed extrapolation tables (e.g. XG0001, XGT0005, XNM0001) are removed; the X and GT tokens no longer appear as table names. This change is forward-only: existing PP_DB_*.db files and old X-prefixed tables from prior runs are not migrated.

Fixed

  • Plotting the Score parameter in the SOP GUI subsection ("Type of parameter to plot" → Plot) no longer raises NotImplementedError: Parameter not parametrised.. The Score output is a computed value (not a perturbed parameter), so get_boundaries now returns a [0.0, init_val] range for it instead of querying the perturbator, and its histogram plots correctly. Non-score parameters are unaffected.
  • The Score-column histogram in the SOP GUI subsection ("Type of parameter to plot" → Plot) no longer draws the brown perturbation-boundary vertical lines; since a score is a computed output rather than a constrained/perturbed parameter it has no boundaries, so only the black init-value line is shown. Non-score parameters still draw their brown boundary lines as before.
  • SOP parameter plotting in the analysis GUI no longer risks crashing from memory exhaustion, since it no longer redundantly rebuilds models and rescores once per selected parameter before plotting.
  • Postprocessing the optimized (GOAT / GT) ensemble no longer crashes with a KeyError (e.g. 'exp_012') during scoring: the experiment-scoring loop now skips experiments whose name has no cached score in mdl.sop.scores, which is the case for the pp_experiments swapped in during postprocessing.
  • When postprocess mode reuses an already-persisted (P, T) rate-coefficient grid (missing_grid is False), CoreRun now calls mdl.rateCoef.recover_rslts() before marking the model KIN, so the cached rate-coefficient results are actually recovered/loaded instead of left unpopulated.

1.0.4 - 2026-07-23

Added

  • Analysis notebook for the ethyl oxidation example included in the example folder.
  • QoL improvements to the experiment class, allowing easy plotting of TimeProfile type experiments in a jupyter notebook.

Fixed

  • Bug in the scoring module that caused the count of active parameters to be incorrectly computed. The issue has been resolved, and the count of active parameters is now independent from the active parameter list used by the perturbation and updated by the sensitivity analysis.

1.0.3 - 2026-07-21

Added

  • Sensitivity analysis can restart with frozen parameters.

Fixed

  • Two-sided derivatives properly skipped for frozen parameters in the linear sensitivity analysis.
  • Minor bug fix in the scoring module to correctly compute the experimental score when species weights are applied.

1.0.2 - 2026-07-20

Added

  • Frozen parameters can now be specified in the input JSON file using the fixed_params key. This allows users to exclude certain parameters from being perturbed during optimization.
  • Working example for ethyl oxidation with frozen parameters included in the example folder.

1.0.1 - 2026-07-14

Added

  • Visualization and export of KMO databases in the database tab of the GUI.
  • Improved score printing for clearer run output.

Fixed

  • Minor print formatting issue.

Changed

  • Unified the package version across pyproject.toml, setup.py, and meta.yaml.

1.0.0 - 2024

Added

  • Initial public release of KiMecO (Kinetic Mechanism Optimizer).