English | 简体中文

Agentic Simulation Lab

Turn GUI-heavy engineering simulation into Agent-orchestrated, reproducible workflows with explicit physics validation and durable evidence.

Quick Start · Simulation Gallery · Documentation · Learning Path

Agentic Simulation Lab makes this chain executable and auditable:

Human physical intent → Coding Agent → reproducible script / CLI → Ansys solver → physics validation → structured result

The Agent discovers and coordinates the work. Ansys software still performs the numerical simulation. Declared physical checks—not a successful process exit—determine whether evidence is PASS, FAIL, BLOCKED, PARTIAL, or NOT_RUN.

Real multi-domain simulation results

This repository contains real multi-domain simulation results. The PNG figures below are deterministic post-processing of qualified, solver-derived numeric evidence—not AI-generated contours or proprietary GUI screenshots.

Mechanics: solved nodal displacement and stress	CFD: transient velocity and pressure fields
Multiphysics: fluid–solid conjugate heat transfer	Acoustics: solver eigenmode pressure slices
Electromagnetics: reconstructed axisymmetric field	DEM: final solved particle configuration
SPH: Lagrangian free-surface snapshots	Phase change: liquid-fraction evolution

Browse all 11 representative figures in the simulation-result provenance inventory or continue to the complete benchmark gallery. Explanatory SVGs remain as a separate orientation layer and are labeled as schematics or domain maps.

See the lab before reading the architecture

These are historical repository results and validation schematics, not guarantees for another version, license, mesh, or machine. The schematics are not fabricated solver contours.

Physics story	Solver / product	Validation basis	Historical evidence
Static cantilever	Mechanical / MAPDL	Euler–Bernoulli tip deflection	PASS — 0.100143 mm vs 0.100000 mm; 0.143% error
Laminar channel	Fluent	Poiseuille profile, pressure drop, mass conservation	PASS — profile L2 error 0.210%; pressure-drop error 0.150%
Conjugate heat transfer	Fluent	Global energy closure and temperature bounds	PASS — 0.900% energy imbalance
Acoustic tube	MAPDL / Mechanical	Quarter-wave resonance	PASS — 86.0 Hz vs 85.81 Hz; 0.221% error
Particle free fall	Rocky	Constant-gravity kinematics	PASS — maximum position error 1.34 µm
SPH dam break	Rocky	Front advance, mass conservation, time history, projection checks	PASS — all declared historical checks passed

Browse the complete visual catalog for all 134 cases across 11 physics domains, including the deliberately visible failures, external blocks, and cases without attributable run evidence.

What problem does it solve?

Traditional simulation work often leaves important state inside GUI clicks, local project files, and one-off scripts. That makes a workflow hard to reproduce, review, delegate to a coding Agent, or reuse for dataset generation. This project gives each case a stable manifest, CLI entry point, solver-local implementation, declared validation logic, structured result contract, and compact evidence record.

It is a laboratory for learning and building trustworthy automation—not a replacement for Ansys products, licenses, qualified engineering review, or numerical judgment.

How it works

1. list and info read solver-independent manifests without importing a proprietary integration.
2. doctor diagnoses the current machine without launching a solver by default.
3. run ... --dry-run resolves the exact command, prerequisites, paths, timeout, and expected result.
4. An explicitly authorized run calls the local solver through a supported API or script interface.
5. The case extracts numerical evidence and applies predeclared analytical, conservation, canonical, dimensional, or physical-trend checks.
6. run.json records process and physics status separately; the manifest's result file remains authoritative.

Read the architecture, validation policy, and Agent operating workflow when you need the full contract.

Quick Start

The core catalog, CLI, dry-runs, tests, and audits do not require Ansys software.

python -m pip install -e ".[dev]"
agentic-sim list
agentic-sim doctor
agentic-sim info cfd fluent-laminar-channel
agentic-sim run cfd --case fluent-laminar-channel --dry-run
agentic-sim validate

For a reproducible project-local environment, use python tools/bootstrap.py --extras dev. Install only the solver extra you need, such as .[mechanical], .[fluent], .[aedt], or .[rocky].

Removing --dry-run is a separate decision: it requires explicit authorization, a compatible official local product, and an available license. Start with the bilingual Quick Start, then follow Run a benchmark.

Explore the lab

If you are…	Start here	What you will find
an engineering or science student	Flagship learning path	Eight staged cases from beam bending to CFD, CHT, acoustics, particles, and datasets
a Mechanical / Fluent / AEDT / Rocky user	Simulation catalog and solver matrix	Reproducible case scripts, product requirements, validation basis, and honest historical status
a Scientific-AI researcher	Dataset guide and dataset tutorial	Parameter sweeps, portable Dataset Contract v1 metadata, safe NPZ loading, checksums, and separate physics provenance
a contributor or tool builder	Development guide and contributing	Manifest schema, result contracts, lazy integrations, static validation, and publication boundaries

The documentation home separates newcomer tutorials from concepts, reference, compliance, reports, and maintainer-only release evidence. It also defines the project's maintainable bilingual policy.

Technical principles

• Local First — no telemetry, automatic upload, or online AI API in the core runtime; solver work runs locally after dependencies are installed.
• API / Script First — supported CLIs, Python APIs, and solver scripting interfaces are the default, not fragile GUI automation.
• Agent / Model Agnostic — any coding Agent that follows the repository contract can inspect and orchestrate the same workflow.
• Physics First — an exit code cannot establish correctness; declared physical evidence must support PASS.
• Reproducible and auditable — project-relative manifests, routed artifacts, bounded subprocesses, explicit provenance, and stable result contracts.
• Honest evidence semantics — known FAIL, BLOCKED, and NOT_RUN cases remain visible because negative evidence is part of scientific credibility.

Current generated counts are in project metrics. Known failures and product/API limitations—including the AEDT electrostatic regression, Turek–Hron FSI, reactive-flow energy accounting, Rocky two-way coupling, and selected SPH modes—are documented in known limitations.

Platform and solver requirements

Python 3.10 or newer is required. The solver-free core and static workflows support Windows, macOS, and CI-configured Linux; local Ansys Student desktop execution is currently documented for compatible Windows installations. Product availability, license terms, model limits, integration versions, and supported transports vary.

See tested environments, solver support matrix, Student product limits, and the platform installation tutorials under docs/tutorials/.

Contributing

Contributions are welcome when they preserve project-relative paths, lazy solver imports, physical acceptance criteria, evidence status, and public-tree privacy. Read CONTRIBUTING.md, use SUPPORT.md for help, and report security issues through SECURITY.md.

After manifest or gallery changes, regenerate and check the public navigation:

python tools/build_catalog.py
python tools/build_project_metrics.py
python tools/build_gallery.py
python tools/build_gallery.py --check
python tools/build_simulation_visuals.py --check
python tools/check_links.py

License, compliance, and disclaimer

This is an independent community project. It is not affiliated with, endorsed by, certified by, or supported by Ansys, Inc. Ansys software and an appropriate license must be obtained separately and used under their applicable terms. The repository does not distribute Ansys software, proprietary solver databases, vendor documentation, logos, or trade dress.

The Apache License 2.0 applies only to rights held by project contributors. It grants no rights in Ansys software, documentation, trademarks, official/vendor assets, or other third-party materials. Solver-derived numbers and project-rendered figures are published as validation evidence without claiming that Ansys licensed them under Apache-2.0. Ansys license types are not interchangeable; users must lawfully obtain the software and license appropriate to their intended use. Engineering results require independent review by qualified practitioners.

Read Ansys usage and compliance, the full disclaimer, and third-party notices.

Ansys and its product names are marks of Ansys, Inc. or its subsidiaries and are used only to identify interoperability targets. Other names may be marks of their respective owners.