Skip to content

Agentic AI and NAVIER AutoResearch

NAVIER-CFD contains an early agentic planning layer and, in release 1.1.0, a persistent Codex-integrated AutoResearch foundation.

The platform separates language-model reasoning from deterministic scientific execution.

flowchart LR
    LM[Language model or Codex]
    PLAN[Plan and explain]
    TOOLS[Explicit NAVIER tools]
    SCI[Deterministic scientific core]
    GOV[Contracts, approvals, budgets]
    EVID[Evidence and provenance]

    LM --> PLAN --> TOOLS --> SCI --> EVID
    GOV --> TOOLS
    GOV --> SCI
    EVID --> LM

Existing planner

The provider-neutral AgentOrchestrator converts a natural-language CFD objective into:

  • an interpreted task;
  • a relevant dataset family;
  • hard-filtered model compatibility;
  • an evidence-aware model shortlist;
  • a benchmark and ablation plan;
  • reasons and cautions.

It can operate offline and does not send data, files, credentials, or private results to an LLM by default.

from navier_cfd.agents import AgentOrchestrator

agent = AgentOrchestrator()
plan = agent.plan(
    "Benchmark sim-to-real cylinder wake forecasting on RealPDEBench, "
    "24 GB GPU, conservation and uncertainty required"
)
print(plan.to_dict())

AutoResearch upgrade

NAVIER AutoResearch adds:

  1. ResearchObjective and ResearchContract;
  2. resource ceilings and deterministic stop policies;
  3. assistant, guided, and bounded modes;
  4. persistent action proposals and approvals;
  5. findings, decisions, and resource accounting;
  6. Codex repository skills;
  7. a local read-only MCP server;
  8. deterministic CFD diagnostics;
  9. FigureLab specifications, audits, renderers, and manifests.

Current agentic capability levels

Level Capability Status
0 Static model and dataset library implemented
1 Natural-language task interpretation implemented
2 Structured model and benchmark planning implemented
3 Tool-using analysis and persistent governance foundation implemented
4 Iterative approved execution and replanning session foundation; execution adapters future
5 Solver-connected active data generation future
6 Autonomous laboratory operation outside current scope

The v1.1.0 release should be described as a scientifically governed AutoResearch foundation, not a complete autonomous scientist.

Current MCP tools

  • dataset catalogue listing;
  • model catalogue listing;
  • deterministic research planning;
  • evidence-aware model recommendation;
  • metric-suite listing;
  • figure-spec auditing.

See AutoResearch MCP tools.

Current Codex skills

  • define a research problem;
  • audit CFD data;
  • run bounded AutoResearch;
  • diagnose CFD results;
  • generate research figures;
  • review scientific validity.

See Codex skills.

Scientific boundary

The language model may:

  • interpret goals;
  • select skills;
  • propose hypotheses;
  • choose deterministic tools;
  • explain calculated results;
  • prepare code and reports.

NAVIER-CFD must:

  • validate schemas;
  • calculate metrics and diagnostics;
  • enforce compatibility;
  • score registered evidence;
  • track provenance;
  • enforce approvals and budgets;
  • apply stop policies;
  • audit figure specifications.

The language model must not invent:

  • units;
  • boundary conditions;
  • field semantics;
  • solver settings;
  • physics metrics;
  • experimental evidence;
  • benchmark results.

Read-only safety boundary

The v1.1.0 MCP server does not automatically expose:

  • training;
  • package installation;
  • downloads;
  • OpenFOAM, MFiX, or DEM execution;
  • Slurm submission;
  • file overwrites;
  • deletion;
  • arbitrary shell commands.

Future execution connectors must use the same research contract, risk classification, approval records, budgets, and provenance requirements.

Complete documentation

Start with NAVIER AutoResearch and Codex integration.