An agent-native open process-engineering runtime, in which models, equations, numerical strategies, validity limits and evidence are explicit objects. Humans and coding agents work against the same scientific contracts. The project owns solver orchestration and process-specific numerical methods, and reuses established differentiation, linear-algebra and optimization infrastructure.
A flowsheet is a validated document, a solve is a job with a typed result, and every converged run
comes with a certificate from an independent verifier that says what it checked and what it does
not claim. Python, the openflowsheet command line, HTTP and MCP run the same operations with the
same results.
Documentation: https://computational-chemical-engineering.github.io/openflowsheet/
Early development. Version 0.3.0 is released (31 October 2026, on PyPI and as a GitHub release).
0.3.0.post1 is the same release with this description and the project links updated; it changes
no code.
| Released in 0.3.0 | Everything 0.2.0 has — the application contract (local, HTTP and MCP), the unit-model library, the planner, the verifier with its certificate or typed failure, the replay bundle, the Peng–Robinson route of the ammonia-synthesis loop, the pinned PyMRM reactor as a unit model, sensitivities, studies and the NLP bridge, surrogate evidence, the trust-region study and the diagnostic web shell — and, new in 0.3.0: countercurrent equilibrium-stage columns with their MESH model, purity, recovery and stage-temperature specifications and one-phase sections (S01, S02, S06, S08, S08b); blind column comparisons with an independent implementation, IDAES, DWSIM and BioSTEAM (S03, S09); IAPWS-95 water (S05); CoolProp's Helmholtz equations of state for 123 pure fluids as an optional extra (S07); the Wilson liquid-mixture method for methanol–ethanol–water with TP and PH flash and the column (S10, S10-C); all of these selectable in flowsheets, with the seven thermal.* units (S11); and cost estimation and LCA as post-solve functionals of a verified state (S12/E01). The supported envelope is the support matrix: unlisted is unsupported |
| Established | Numerical verification against independent references, for example the P01 oracle against 20-digit reference values (7.1e-15 worst relative deviation); agreement with IDAES, DWSIM and BioSTEAM on shared column cases, which share the equations and are therefore not validation |
| Validation against data | The Wilson method is within its pre-registered bands against measured vapour–liquid equilibrium of methanol–ethanol–water (ternary at 101.3 kPa; binaries with water at 323–333 K); methanol's liquid enthalpy and heat capacity are not within band against CoolProp. The C1 loop and its reactor results are not empirically validated |
| Not established | Optimality evidence; human numerical or process-modeling review — review.numerical and review.process_model are pending in every evidence manifest and no agent may set them |
What 0.3.0 claims, and what it is not (Linux only, no global optimality, no liquid–liquid split, no rate-based columns, among others), is in the changelog.
OpenFlowsheet needs Python 3.13 or newer and runs on Linux (x86-64 and aarch64 are tested).
python -m pip install openflowsheet| Extra | Install | Adds |
|---|---|---|
coolprop |
pip install "openflowsheet[coolprop]" |
The CoolProp pure-fluid basis coolprop-heos-v1 |
server |
pip install "openflowsheet[server]" |
The HTTP server with the optional web shell, and the MCP server |
lca |
pip install "openflowsheet[lca]" |
The offline Brightway adapter that builds LCA unit-score tables |
nlp |
see the installation guide | The general NLP bridge (Pyomo and Ipopt); needs a conda-forge environment |
Solve a once-through Rankine cycle on IAPWS-95 water from a ready-made change set:
curl -sLO https://raw.githubusercontent.com/computational-chemical-engineering/openflowsheet/v0.3.0/examples/rankine-cycle.changeset.jsonimport json
from openflowsheet.application.local import LocalApplication
from openflowsheet.application.types import Change
with open("rankine-cycle.changeset.json") as handle:
change_set = json.load(handle)
change, expected_revision, key = Change.from_change_set(change_set)
with LocalApplication.create("plant", exist_ok=True) as app: # created, or opened on a re-run
committed = app.commit_change(change, expected_revision, key)
report = app.validate(committed.revision_id, "simulation")
print(report.status) # READY_FOR_SIMULATION
run = app.solve(committed.revision_id, "default")
print(run.outcome, run.verification_status) # CONVERGED VERIFIED
run.raise_for_outcome() # says why, if the solve failed
bundle = next(o.artifact_id for o in run.outputs if o.kind == "replay_bundle")
state = json.loads(app.artifact_bytes(f"{bundle}/solution-state.json"))["variables"]
print(state["T.W"], state["B.Q"]) # turbine work, boiler duty (W)The same steps from the command line:
openflowsheet project init plant
openflowsheet api --project plant commit_change --json-file rankine-cycle.changeset.json
openflowsheet api --project plant solve --json '{"revision_id": "rev-000001", "policy_id": "default"}'The quickstart reads the certificate, and the tutorials build a flowsheet step by step.
- Documentation: https://computational-chemical-engineering.github.io/openflowsheet/
- Repository: https://github.com/computational-chemical-engineering/openflowsheet
- Changelog: https://github.com/computational-chemical-engineering/openflowsheet/blob/main/CHANGELOG.md
- Releases: https://github.com/computational-chemical-engineering/openflowsheet/releases
- Issues: https://github.com/computational-chemical-engineering/openflowsheet/issues
- Support matrix: https://github.com/computational-chemical-engineering/openflowsheet/blob/main/docs/support-matrix.md
- Limitations: https://computational-chemical-engineering.github.io/openflowsheet/guide/limitations/
This software uses CasADi 3.8.0, which is licensed under the GNU Lesser General Public License, version 3 or later (LGPL-3.0-or-later). CasADi is a separate work, obtained by the installer from PyPI as a pinned dependency; this project distributes none of its bytes. It is used through its Python API only — the interpreter loads the library at run time, nothing is compiled against its headers or linked against it at build time, and a user may substitute their own build of CasADi 3.8.0 by replacing the installed package.
The full licence text ships inside the CasADi distribution at casadi/LICENSE/LICENSE.txt in the
installed package, together with notices for its bundled components at casadi/include/licenses/.
An exact inventory of what that distribution contains — every binary with its SHA-256, the bundled
plugin inventory including the bytes that cannot run, and the notices — is
docs/p03-binary-audit.md,
with the machine-readable record in spikes/p03/results/. The terms on which this project may and
may not redistribute any of it are
docs/adr/0006-distribution-data-rights.md.
Changing the CasADi pin re-runs that audit before the change merges.
Further attributions are in
NOTICE.
The core licence is Apache-2.0, declared in pyproject.toml and stated in
LICENSE and
NOTICE. The
maintainer confirmed the project's rights to distribute its contributions under it on 2026-09-23
(120acb4). Why Apache-2.0 rather than MIT is recorded in the
decision register
R-005.
git clone https://github.com/computational-chemical-engineering/openflowsheet.git
cd openflowsheet
make venv # .venv with the project and its dev extra, from the pinned lock
make check # ruff, ruff format, mypy --strict, pytest — the gate; a failed gate stays failedThe architectural authority is the
blueprint,
the execution authority the
implementation plan,
and the normative decisions are the
ADRs, indexed by the
decision register.
Coding agents follow
CLAUDE.md.
What each package has and has not shown is in evidence/<ID>/<commit>/manifest.json, the current
position in
docs/progress.md,
and how a release is made in
docs/RELEASING.md.
The diagnostic web shell (openflowsheet serve-http --ui, with the server extra) is described in
docs/web-shell.md.