A 2,000 L run commits weeks of media, a suite, and a slot on the schedule. Stillwater models the reactor first — and returns every prediction with the interval around it, because a titer number without an envelope is not a decision input.
Fed-batch dynamics integrated with RK4, Monod growth kinetics with substrate inhibition, Luedeking-Piret product formation, and death kinetics driven by ammonia, lactate and pH. Mass balance is closed and checked, so the trajectory is physically coherent rather than curve-fitted.
A Gaussian-process surrogate trained on the parametric model returns a posterior, not a point. You get the mean, the standard deviation, and the P10/P50/P90 envelope — which is what tells you whether a setpoint is comfortably inside spec or one bad lot away from an excursion.
Two-parameter parametric sweeps and Monte Carlo sampling across the operating envelope, classified into proven-acceptable and at-risk regions, exported as an ICH Q8 design-space artefact with the e-signature attached.
IQ, OQ and PQ protocol documents generated as PDFs against the actual deployment — installed versions, database role grants, service units, network posture. Templates for your QA team to review and execute, not a certificate.
Send an anonymised historical batch. Stillwater predicts the outcome without seeing the result, and only then reveals it. That is the only honest way to answer "does this twin work on my process" before anyone signs anything.
SHA-256 hash-chained audit log, serialised under a PostgreSQL advisory lock so concurrent writes cannot fork the chain. UPDATE, DELETE and TRUNCATE are revoked from the application role at the database level, with a trigger as defence in depth.
Some of this platform is first-principles calculation. Some of it is correlation and parameter fitting. Both are legitimate; conflating them is not. This table is the honest version, and the software logs its own provenance at startup so a deployment cannot quietly drift from it.
This deployment is not CFD-informed. Stillwater can consume per-vessel CFD transfer-function datasets when a site supplies them. Where none are present, scale-effect corrections come from published engineering correlations instead — which is the case by default. The service logs which one it is using at startup, and tags every result with its source, so the distinction survives contact with a report.
This is not genome-scale flux balance analysis. The CHO and E. coli parameter sets are keyed to the published iCHO1766 and iJO1366 reconstructions and cite them, but the solver does not build a stoichiometric matrix or run linear programming over thousands of reactions. It selects among pre-computed flux distributions and scales them. Useful, fast, and bounded — and not the thing a reader might assume from the model names.
Deployed as an on-premise appliance for air-gapped CDMO environments, with OPC-UA ingest from the historian and no dependency on outbound connectivity.
Open Stillwater