Local verification

Run numerical checks locally from the repository root. GitHub Actions builds this website, checks documentation links and runs Python tooling tests; it does not execute MATLAB or establish numerical correctness. A green documentation check is separate from the MATLAB verification record below.

Choose the checks for your installation

H2-store unit and numerical tests require the pinned MRST dependency and do not need PHREEQC, internet downloads, plots or Parallel Computing Toolbox during the test run. Initialize the submodule beforehand:

addpath(fullfile(pwd, 'tests', 'h2store'));
results = runH2StoreTests();

The runner restores the original MATLAB path and working folder and requires every test to pass without skips. It can also accept an existing pinned MRST checkout as its argument.

For biochemistry tests, first follow Installation and initialize H2sim:

startupH2sim('phreeqc', true)
mrstModule add test-suite
checks = runtests(fullfile(pwd, 'tests', 'biochemistry'));
assertSuccess(checks);

Some tests require a PHREEQC engine or optional parallel support. Inspect incomplete results separately; a skipped test does not validate its backend. The fully configured verification record below contains no incomplete tests.

Complete-cycle and serialized-restart regressions require working PHREEQC:

addpath(fullfile(pwd, 'tests', 'cycles'));
cases = runStorageCycleRegressions();

This runner checks two four-cell cases: black-oil dissolution/vaporization and reactive biochemistry with PHREEQC. Each has six steps spanning injection, shut-in and production. A MAT-file checkpoint within shut-in is loaded into a fresh model and solver before resuming production. The PHREEQC check also uses a fresh engine session. Physical-state bounds, well control changes, hydrogen inventory accounting for the black-oil case, reaction consumption, restart agreement and the reviewed saved reference must pass. Missing dependencies or failed comparisons raise an error; the runner never regenerates references.

See the cycle regression guide for comparison bounds, reference provenance and diagnostic inspection commands. Run test folders explicitly after activation; they are intentionally outside the runtime module path.

Recorded local run

This is a dated record, not an automatically refreshed badge:

Verification on 29 October 2026

Item

Recorded environment or result

Source

EOS fix d3dbf3f and reviewed reference fc6e2fa, merged in a8b56fc

MATLAB / platform

R2026a Update 5, Windows MATLAB launched from WSL

MRST

Pinned commit 8e218b0786cd4651814553b0f9ebd6844eaabcb2

PhreeqcMatlab / IPhreeqc

Pinned commit 20662bf8fab5fe2a9bcaa7b60bc7f1ae5337733f; IPhreeqc 3.8.6

Biochemistry

121 of 121 tests passed; zero failed or incomplete

H2-store

217 of 217 tests passed

Complete-cycle regressions

Both model cases passed, including fresh-model serialized restarts

Documentation and tooling

Strict Sphinx build, local HTML links/images and seven Python tooling tests passed

The physical HS restart discrepancy was traced to an EOS trace-composition floor comparison and corrected without widening regression bounds. The maximum HS tracer difference in the reactive cycle decreased from 7.170710469e-8 to 1.394717675e-14 mol/m^3. The changed uninterrupted reference was reviewed and its old/new hashes and numerical differences recorded.

The revised coarse-grid introductory aquifer command was also executed in this environment with fresh output: 33 cells, 179 accepted control steps and one 220-day simulated build-up/storage cycle. Its explicit plotting call passed. This manual onboarding check is separate from the 338 MATLAB test results and uses a different grid and cycle count from the archived five-cycle figures.

Optional-startup verification

On 29 October 2026, startup change 4b86044 was checked in the same Windows MATLAB R2026a Update 5 environment, using archived pinned MRST and PhreeqcMatlab source files rather than locally edited dependency checkouts:

  • With no PhreeqcMatlab directory, default startup completed three black-oil phase-transfer cases and a two-step compositional simulation. The chemistry checkout remained absent. A requested native interface reported H2sim:PhreeqcNotInitialized with explicit setup instructions.

  • Default and explicit-false startup passed repeated-call and changed-working- directory checks without initializing chemistry.

  • Explicit chemistry startup initialized the real IPhreeqc engine; both complete- cycle/reference/restart cases passed. Repeating core startup preserved the already initialized chemistry paths.

  • All 126 biochemistry tests and 217 H2-store tests passed, with no failed or incomplete biochemistry tests. The five additional biochemistry tests cover startup options, explicit initialization and dependency errors.

  • All 13 Python tooling tests passed. The six launcher checks exercise core startup without a chemistry checkout, opt-in Linux preloading, Windows routing, path quoting, argument errors and MATLAB exit-status propagation using fake executables. They do not execute native Linux MATLAB.

The focused startup tests can also run without a chemistry checkout:

startupH2sim
checks = runtests(fullfile(pwd, 'tests', 'biochemistry', ...
    'H2simPhreeqcStartupTest.m'));
assertSuccess(checks);

No numerical references, chemistry parameters or solver acceptance bounds were changed for optional startup. The earlier dated record remains above for the EOS restart correction and reference provenance.

Platform and scientific limits

The stated minimum for core workflows is MATLAB R2021a. This record does not establish results on every MATLAB release. Native Linux PHREEQC coupling is implemented, but a complete native-Linux MATLAB regression run is not recorded here. Native-interface chemistry tested with Windows DLLs is not a substitute for an end-to-end Linux run. macOS sequential PHREEQC coupling is unsupported.

The hybrid PE diagnostic remains history-sensitive and is reported separately from physical restart comparisons. The existing checks require finite PE and retain it in the uninterrupted reference; they do not require PE restart equality. See Code reference for the chemistry ownership and restart details.

These are unit checks and small numerical cases, not full paper-duration simulations, exhaustive phase-transition coverage, or an independent reactive elemental-conservation audit. Reference agreement detects numerical change; it does not prove that every physical model is correct. Papers and archived figures may use different grids, schedules and parameter sets.

Record a new environment

When reporting another validation run, record the MATLAB version and platform, H2sim commit, pinned dependency revisions, backend and library version, and all passed/failed/incomplete counts. From the checkout, capture:

git rev-parse HEAD
git submodule status

version
computer
which BiochemistryModel

Keep earlier results as dated records. A reference update must explain the numerical change and retain a before/after review; do not regenerate a reference or relax its bounds simply to make a failing test pass.