Skip to content

epcsaft 1.3.1

Latest

Choose a tag to compare

@tannerpolley tannerpolley released this 17 Nov 06:03
· 50 commits to main since this release

Scientific source distribution for ePC-SAFT Engine 1.3.1, cut from candidate 1b46e57ce8f78b79891b5fe5293917c58fa0f671.

Documentation: https://tannerpolley.github.io/epcsaft/

Changes since 1.3.0

  • CppAD thread-slot cleanup avoids promise destruction during Windows library initialization while preserving safe thread-exit reuse.
  • Windows bindings can link the native active-parameter value accessor, restoring the extension build.
  • Zero-loading reactive states can now be searched by the Full stability search and used with the stability penalty when a reacting inventory is absent from the feed. Generated trial phases exclude species proved absent by the conserved inventory constraints, instead of requiring a positive amount of an unavailable species. This restores the search on the admissible composition face; it does not establish global stability.
  • More equilibrium specifications: neutral fixed-volume TV and UV flashes; charged and source-referenced PH and PS flashes; apparent equilibrium heat capacity for a fixed-pressure enthalpy specification; reaction-relaxed stability checks. Searches remain finite and do not prove global stability.
  • Additional properties and derivative actions: second and third virial coefficients, native state arrays, fugacity coefficients by contribution, and referenced second-order actions in EOS-parameter and reaction-law directions.
  • Broader process-model declarations: ASL records accept active EOS parameters, temperature and pressure controls with reaction laws, system enthalpy and thermochemistry, liquid–vapor pairs, and declared-present pure solids. Charged-species pressure-property records carry exact derivatives.
  • Parameter estimation can fit a solid's equilibrium-constant law to solubility observations and impose tangent-plane stability requirements. Reaction enthalpy-shift coordinates carry through fit sequences and minimax estimation.
  • Pathway-weighted reactive cross-association is available with exact derivatives. Dilute-limit arithmetic, trace-ion splits, pure saturation recovery and continuation admission have been corrected. Numerical failures retain their cause and finite diagnostics.
  • Parameter records can add cation-pair interactions. The HCl packet figiel-2025-chloride-mixture-refit/4 includes fitted H–Na and H–Mg interactions. At 298.15 K and 1 bar, the lowest-acid measured checks are 0.1 mol/kg HCl in NaCl at ionic strengths 4, 5 and 6 mol/kg (acid contributes 1.7–2.5% of ionic strength). Above 6 mol/kg, checks extend to 7 mol/kg only for acid fractions of at least 10%, in HCl–NaCl and HCl–LiCl. H–Mg is checked to ionic strength 5 mol/kg, H–Ca to 3 mol/kg; H–K is unmeasured. Trace-acid states, acidic states above 7 mol/kg and acidic brines with multiple salt cations have no measured comparison. HCl–water at ionic strength 7–16 mol/kg remains report-only, with worst discrepancy 0.07433 decade. Inherited chloride-packet limits and the absence of propagated parameter uncertainty remain.
  • Packet domains can state composition bounds and report applicability on request. Temperature-law admission now checks the declared temperature interval consistently. Density diagnostics share their native defaults.
  • The public repository now directly owns Engine source. Private research inputs and analysis trees stay outside this distribution. The two MIT FeOs Gross–Sadowski packets remain installed in the conda package; omitted non-commercial FeOs packets can be restored from pinned upstream bytes with bash data/tools/restore_feos.sh.

Verification on the exact archive bytes before publication

  • The documented locked setup and source installation pass from the extracted archive outside a Git checkout and load 1.3.1. The README NaCl example prints exactly the documented activity and osmotic coefficients.
  • linux-64 / CPython 3.13: the conda package built from these bytes passes its import, dependency and numerical tests and includes exactly the two MIT FeOs Gross–Sadowski packets. The installed test gives pressure 1695742.8953910787 Pa and temperature action 39302.24177938896 Pa/K, with largest central residual 8.88e-16 and action linear residual 3.87e-17.
  • The corpus comparisons were run at candidate 3c2046a6; the final corrections change Windows symbol export and thread-slot cleanup without changing thermodynamic formulas. These extended comparisons have not been rerun on the final archive. With OpenBLAS and MUMPS, the installed and development builds agree with zero corpus comparison failures: Cameretti 1800/1800, Gross 721/724, neutral sweep 63/70, matrix 42/42, replay 4385 rows and MEA flash family 73/73. The known unavailable states are retained on both sides.
  • Key solves and the full fit have zero comparison failures. Largest scaled state difference is 1.0036538265939643e-11 and largest scaled full-fit difference is 6.443695845532716e-12, both below the existing 1e-10 comparison tolerance. Full-fit costs are 31.12372699861994 (development) and 31.123726998610728 (installed), an absolute difference of 9.212186569129699e-12. Both fits take 14 iterations.
  • All inner and outer source-set SHA-256 checks pass. Native package libraries contain no home-directory or local build paths; local /tmp paths remain in Python bytecode, installation metadata and recipe metadata.
  • The full documentation build passed at candidate 3c2046a6: all twelve example notebooks are executed, both Sphinx warning logs are empty, and all 37 doctests pass. The settler example now supplies fixed temperature and pressure to the ASL callback; Sphinx displays Mermaid diagram source as plain text while GitHub keeps its rendered diagram.
  • The standard host fast checks pass. An additional run forcing MUMPS fails one native endpoint-observation precheck; the identical executable passes with the ordinary host solver. This solver-dependent assertion also fails at the 1.3.0 source and is a separate check defect; it does not show a new equilibrium failure. The package tests, documentation, and installed/development numerical comparisons above pass with MUMPS.
  • win-64 / CPython 3.13: the conda package built locally from these archive bytes, with MSVC v142 14.29 clang-cl and Boost headers 1.89.0 h57928b3_3 from the source set, passes its import and run tests and imports 1.3.1 in a fresh environment. The installed test gives pressure 1695742.8953910803 Pa and temperature action 39302.241779389 Pa/K. These differ from linux-64 by 1.6e-9 Pa and 3.6e-11 Pa/K, within the 1e-8 package tolerance. Four threads evaluating the same pure-ethane state, 300 K and 10,000 mol/m³, agree at 4316185.179414262 Pa, and the process exits normally. The Windows native libraries embed seven Boost.Multiprecision exception source-location header paths from the build tree. They contain no home directory or username, and 1.3.0's Windows libraries carried six of the same header paths from its CI build.

The corresponding-source set includes this archive, Ipopt 3.11.9 with its patch, embedded dependency sources, Boost 1.89.0 headers, original notices and the exact recipe. Coin-HSL is not distributed.

Packaging comparisons are numerical verification of how model predictions carry through packaging. They do not establish physical validation. The HCl range above describes the packet's retained measured comparisons and their limits. Other platforms and Python versions are not qualified by the Linux packaging checks.