optimize neqsim process model plant

Loaded skills

Loaded skills: neqsim-agentic-process-optimization, neqsim-optimization-and-doe, neqsim-platform-modeling, neqsim-api-patterns, neqsim-notebook-patterns, neqsim-professional-reporting

ALWAYS read these skills before proceeding:

• .github/skills/neqsim-agentic-process-optimization/SKILL.md — the core recipe: decision space, robust trial evaluation, objective/constraint extraction, multi-year sweeps, and the NeqSim-update workflow
• .github/skills/neqsim-optimization-and-doe/SKILL.md — which optimizer algorithm to pick (SQP, PSO, BatchStudy, ProcessSimulationEvaluator bridge)
• .github/skills/neqsim-platform-modeling/SKILL.md — how the multi-area ProcessModel is built (separation trains, recompression, export, recycles)
• .github/skills/neqsim-api-patterns/SKILL.md — fluid creation, flash, equipment accessors
• .github/skills/neqsim-notebook-patterns/SKILL.md — when delivering a notebook
• .github/skills/neqsim-professional-reporting/SKILL.md — when delivering a report

Operating Principles

1. Optimize the live plant, not a toy. The flowsheet is a ProcessModel with named ProcessSystem areas and recycles. Read it, do not rebuild it from scratch.
2. Read the decision space first. ProcessAutomation.getAdjustableParameters() gives bounded knobs. Supply real [lo, hi] for any UNBOUNDED_THRESHOLD (1e9) bound. This only works if the model was built with the optimization data basis — line sizes + manifold sizes for hydraulics, valve/choke Cv, compressor/pump maps + speeds, separator dimensions + design K, and equipment design limits (rated power, surge margin, NPSH, erosional velocity, design P/T, MAWP). Without these, the decision space is empty and capacity constraints never fire. Gather/gate the basis with enterprise-process-model-build-verify (target_fidelity="optimization_ready"). Enable capacity limits for every equipment type (separators, pumps, valves, pipelines, heaters/coolers, HX, manifolds) with ProcessAutomation.enableCapacityConstraints(), and for a pure max-throughput study use the native findMaxThroughputJson(feedAddresses, min, max, unit, utilizationLimit) (bisects total feed to the first binding unit). Routing is a first-class knob: splitters expose bounded splitFactor_i (0-1) INPUTs that getAdjustableParameters() surfaces automatically.
3. Never optimize an ADJUSTER-controlled knob — the model already solves it.
4. Gate every trial. Use runUntilConverged + getRunStatus().success + getConvergenceReportJson(). A non-converged or failed-unit trial returns a large penalty, never a misleading objective and never a crash.
5. Score from real equipment. Compression power and surge margins come from Compressor.getOperatingPoint(); RVP comes from Standard_ASTM_D6377 RvpResult. For a one-call product-quality observable use ProcessAutomation.getProductQualityJson(address) (RVP/TVP + gas cricondenbar/cricondentherm on a cloned fluid, never throws). Test compressor margins for NaN before comparing. Pick the compressor control mode deliberately: solve-speed (fixed discharge P, speed/power are outputs) for spec/capacity/max-throughput studies against fixed pressure boundaries; predictive (fixed speed → computed discharge P) when shaft speed is a decision variable (see neqsim-agentic-process-optimization §5).
6. Soft constraints. Express RVP ≤ spec and surge ≥ floor as penalties so gradient-free optimizers degrade gracefully.
7. Rebuild per year. Feed composition/rate change with the year, so the whole heat/mass balance changes — optimize on a fresh build, not a stale one.
8. Parallel only with deep copies. Use ProcessSystem.copy() / BatchStudy / MonteCarloSimulator for fixed-year screening; copies are independent. For scoring a list of candidate setpoint maps in one call, ProcessAutomation.evaluateBatchJson( candidates, unit, readbacks, maxParallel) runs each candidate on its own ProcessSystem.copy() thread (parallel, live model untouched) and returns per-candidate convergence detail — ideal for SciPy/BoTorch/GA populations.
9. Pick up new NeqSim functionality via devtools (target/classes) or by repackaging the shaded JAR into the pip neqsim (see skill §8). Verify the new methods are callable before relying on them.
10. Honest about gaps. NeqSim has no Bayesian optimization / MINLP / LHS — bridge to SciPy/Pyomo/BoTorch via ProcessSimulationEvaluator when needed.