FluentAI – Autonomous CFD Simulation Assistant Powered by RAG, LLMs, and Rule-Based Reasoning

Created by Duc Phi Ngo (Mr. Bill) — a senior CFD/FEA engineer and AI/ML software developer, FluentAI bridges physics-based simulation with intelligent automation through structured reasoning and multimodal control.

---

Project Vision

FluentAI redefines how engineers interact with simulation platforms like Ansys Fluent, STAR-CCM+, and OpenFOAM. By replacing GUI-based workflows with natural language interaction, FluentAI empowers users to control and automate simulation tasks using:

• Retrieval-Augmented Generation (RAG) for context-aware knowledge recall
• LLM-driven reasoning (LLaMA 3 / GPT) for understanding simulation logic
• Rule-based action planning to ensure reliable and deterministic execution

FluentAI is more than a chatbot — it's a decision-making agent capable of understanding, planning, and executing complete CFD workflows from voice or text input.

---

Why It Matters

In modern simulation workflows, engineers lose productivity to:

• Repetitive GUI operations across platforms
• Manual setup of simulation parameters and post-processing
• Lack of intelligent interfaces for engineering software

FluentAI eliminates these pain points by introducing:

• A multi-intent voice/text interface
• A domain-aware LLM engine that understands simulation language
• A deterministic automation layer built on PyFluent and structured decision rules

---

Key Capabilities

Multimodal user input

• Accepts both voice (via Google Speech-to-Text) and typed commands
• Classifies intent as question, knowledge query, or action plan

Retrieval-Augmented Generation (RAG)

• Uses FAISS/ChromaDB to retrieve relevant engineering documents (YAML, PDFs, Fluent settings)
• Employs LangChain for structured RAG integration

AI and natural language processing

• Powered by LLaMA 3 (or GPT-4) for contextual understanding and planning
• Generates JSON-formatted action plans using deterministic prompt logic and condition-action rules

Deterministic simulation control

• Executes structured commands through PyFluent API:
  • Inlet velocity/temperature settings
  • Turbulence model configuration
  • Solver iteration count
  • Post-processing (contours, plane slicing)
• Currently supports steady-state flow simulations

Modular CFD automation

• Uses a declarative input.yaml to define geometry, boundary conditions, solver settings, and outputs
• Translates LLM decisions into safe, auditable Fluent operations

Software engineering stack

• Python 3.10+, modular architecture
• CLI + voice support for flexibility in demos and real use
• Logs, temp files, and YAML for traceability and reproducibility

---

Fluentai Architecture Overview

User (Voice/Text)
    ↓
[1] Whisper (Speech-to-Text)
    ↓
[2] RAG Engine
    - Retrieves relevant CFD context from Vector Store (FAISS/Chroma)
    - Combines with command to form enriched prompt
    ↓
[3] LLM (LLaMA 3 / GPT-4)
    - Interprets enriched prompt
    - Generates YAML modifications or Fluent instructions
    ↓
[4] YAML Editor
    - Updates input.yaml dynamically
    ↓
[5] PyFluent Automation
    - Runs geometry → mesh → setup → solve → post-process
    ↓
[6] Output
    - Contour plots, reports
    - Voice status updates via gTTS

Voice Output

Currently powered by Google Text-to-Speech (GTTS) for demo simplicity. Will be upgraded to Bark or ElevenLabs in future versions for more natural, production-ready voice output

---

How to Run FluentAI

1. Install dependencies

Make sure you have Python 3.10+ installed. Then run:

pip install -r requirements.txt

Optional (recommended): Create and activate a virtual environment

python -m venv pyfluent-env
pyfluent-env\Scripts\activate  # On Windows only

2. Configure Simulation in input.yaml Edit the input.yaml file to define:

• Geometry path
• Boundary conditions
• Iterations, solver model
• Post-processing outputs

geometry_file: geometry/Static Mixer geometry.dsco
iterations: 50
velocity_inlets:
  - inlet_name: velocity-inlet-1
    velocity: 3.0

3. Run

py ./launch_fluentai.py

Run steady-state simulation for 50 iterations and set inlet temperature to 300 K

4. View Results After a run, check:

• output_*/: for contour images