Studies and optimisation

The studies layer (openflowsheet.studies) works on a compiled problem's pinned inputs. It is a library layer: apart from the surrogate study and revision optimisation, which run as jobs (submit_job with the surrogate_study and optimize operations), it is not an operation of the application contract, and its scope is the registered flowsheets it was built and tested on — the SYN-001 flowsheets and the ammonia-synthesis (C1) loop.

Capability	What it gives	Scope
Parametric sensitivities	exact first-order sensitivities of outputs to pinned inputs, from the parametric twin of the compiled problem and one factorisation, issued only at a certified regular root	registered SYN-001 flowsheets
Sweeps	a parameter varied over values, each point solved and certified, with its sensitivity	as above
Parameter estimation	least-squares estimates of pinned inputs, using the exact sensitivities	as above
NLP bridge	the flowsheet as a grey box inside Pyomo, solved by Ipopt	needs the nlp extra in its audited environment
Revision optimisation	a stored revision optimised over its own specifications, every candidate re-solved and checked (V1–V7), a record that replays; see Optimising a revision	revision_eo revisions; the audited nlp environment on Linux x86-64
Surrogates	surrogate models of an expensive unit with their evidence (SurrogateManifest)	the C1 reactor
Trust-region study	a trust-region refinement of the C1 loop with surrogate and full-fidelity reactor	the C1 loop
Cost and impact outputs	cost and life-cycle impact as study outputs and NLP objectives (assessment:<output_id>), with exact derivatives; see Cost estimation and LCA	sensitivities on any solved revision; the NLP on SYN-001

A sensitivity is never computed by finite differences: differencing is a test oracle only. Studies emit Study and OptimizationReport documents with their schemas.

No optimality claim

Optimality evidence is not established for any study. An optimisation result reports what the solver returned and how it was checked; it does not claim a global or even a certified local optimum. KKT_POINT_VERIFIED is numerically verified local stationarity of the held-regime problem (what verified means).

The nlp extra

The NLP bridge needs Pyomo, cyipopt and Ipopt with libpynumero_ASL. PyPI has no cyipopt wheel for CPython 3.13, so the extra installs only into a conda-forge environment; the repository's scripts/build-m03-ipopt-env.sh builds the audited one, whose binaries and licences are inventoried in docs/m03-ipopt-audit.md. Tests that need it are marked nlp and deselected by default.

A what-if loop through the application

For flowsheets outside the studies layer's scope, a sweep is a loop of ordinary solves: edit one specification with commit_change, solve, read the state. Every point is a revision in the project's history with its own certificate. The what-if example does this for a Rankine cycle.

The normative texts are ADR 0031 (sensitivities and studies) and ADR 0032 (the NLP adapter), linked from Design.