Agentic Process Simulation — Capability Rating

Date: 2026-05-30 Scope: neqsim.process.{automation, dynamics, synthesis, processmodel} and neqsim.util.agentic. Baseline: pre-v1 (before commits 6eed8f50b4, 893c7ccb70, baf2a01f93, f10b0df025).

Overall: 7.5 / 10

Production-ready for read/write and steady-state agents; dynamics and synthesis are MVP.

Dimensional scorecard

#	Dimension	Score	State
1	Discovery & introspection	9/10	`describe()`, `getTopology()`, `getNeighbors()`, `snapshot()`, `getVariableList()`, `getAreaList()` all return JSON with `SCHEMA_VERSION=1.0`. Agents navigate any flowsheet without Java reflection.
2	Read path (typed + structured)	9/10	`getVariableValue`, `getStructured` (composition/k-values), `getValues` batch, `validateAddress`, `getAllowedUnits`, fuzzy auto-correction.
3	Write path (typed + transactional)	9/10	v1: `WriteValidatorRegistry` + `DefaultWriteValidators` (range, type, read-only, unit conversion). `setValuesWithRollback` is true transactional — reverts all on first failure. `getWriteHistory()` audit log.
4	Self-healing & learning	8/10	`AutomationDiagnostics`: fuzzy name match (edit-dist ≤2), learned-correction cache, physical-bounds validation, thread-safe history.
5	Steady-state run loop	8/10	`runIfDirty()`, `setVariableValueAndRun()`, dirty flag, cached facade across turns. Recycle/adjuster orchestration solid.
6	Dynamic simulation	6/10	v1: pluggable `IntegratorStrategy` (Explicit Euler + BDF-1 with FD Jacobian + fallback flag) wired into `ProcessSystem.runTransient` and `ProcessModel.runTransient`. `EventScheduler` for ESD/setpoint events. 3 new measurement devices.
7	Flowsheet synthesis	5/10	v1: `SeparationDuty` + `FlowsheetSynthesisEngine` + `FlowsheetProposal` rank topologies on TAC / recovery / energy. MVP — separation duty only.
8	Quality gates & validation	8/10	`TaskResultValidator` (Java + Python mirror), `SimulationQualityGate`, `AgentBenchmarkSuite`, `AgentFeedbackCollector`, `AgentSession`. CI gate in `task_quality_gate.yml`.
9	State lifecycle & reproducibility	8/10	`ProcessSystemState` / `ProcessModelState` — JSON snapshot, save/load, `validate()`, `compare()` with `ModelDiff`, gzip transfer. Git-diffable.
10	Skill / agent ecosystem	9/10	45+ skills, 30+ agents, semantic skill search, CI-enforced agent↔skill mapping.
11	Multi-area orchestration	7/10	`ProcessModel` with area-qualified addresses (`Area::Unit.property`), scheduler/integrator propagation, plant-wide `getAutomation()`.
12	Documentation & onboarding	8/10	`AGENTS.md`, `copilot-instructions.md`, `CHANGELOG_AGENT_NOTES.md`, `TASK_LOG.md`, dynamic-simulation skill all current.

Before vs. after v1

Layer	Pre-v1	Post-v1
Read agents	8	9
Write agents (mutating control)	5	9
Dynamic / event-driven agents	3	6
Design / synthesis agents	2	5
Reproducibility / governance	7	8
Weighted overall	5.7	7.5

Top remaining gaps (priority order)

~~Adaptive / higher-order integrator~~ — ADDRESSED in v1.1 by RK4Integrator (fixed-step classical RK4) and AdaptiveRK45Integrator (Cash–Karp 5(4) with adaptive sub-stepping and tolerance control). True higher-order BDF (BDF-2..5) and Radau IIA still open for very stiff CCS / refrigeration trains.
~~Sensitivity & gradient API~~ — ADDRESSED in v1.1 by neqsim.process.automation.SensitivityAnalyzer (finite-difference gradients/Jacobians through ProcessAutomation, CENTRAL/FORWARD modes, JSON with SCHEMA_VERSION="1.0"). Analytic / AD-based gradients still future work.
~~Synthesis breadth~~ — PARTIALLY ADDRESSED in v1.1 by CompressionDuty + FlowsheetSynthesisEngine.proposeAndBuildCompression (multi-stage compression with inter/after-coolers). HEN synthesis (pair with PinchAnalysis) and reactor-separation-recycle superstructures still open.
Cross-session learned corrections — persist AutomationDiagnostics correction cache to disk so a new agent invocation starts already calibrated to the model.
Convergence-aware retry policy — runIfDirty() should classify convergence failures (CONVERGENCE_FAILURE exists in taxonomy but no auto-recovery: damping, reset to last good snapshot, switch solver).
DAE / index-1 algebraic constraints — IntegratorStrategy currently treats the system as pure ODE; controllers with algebraic loops would benefit from a true DAE solver.
Parallel multi-area execution — ProcessModel.runTransient iterates areas sequentially; independent areas could run in parallel threads.
Signed / versioned state snapshots — ProcessModelState needs schema-version migration and optional digital signature for audit trails.

Verdict

NeqSim is now clearly the strongest open-source process-simulation substrate for autonomous agents. The v1 commits closed the two biggest credibility gaps (untyped writes without rollback, standalone dynamics infrastructure not wired into the live transient loop). The v1.1 follow-up closed the three top “depth” gaps: more integrator options (RK4Integrator, AdaptiveRK45Integrator), gradient access (SensitivityAnalyzer), and broader synthesis coverage (multi-stage CompressionDuty). Remaining gaps are now genuinely deep-domain work (true DAE solver, HEN synthesis, parallel execution).