TUM-EDA/Flui3d-Chat-Qwen3-Reasoning

Flui3d Chat Model Qwen 3 Reasoning

Model Description

This model is a Fine-tuned version of Qwen 3 designed for microfluidic chip design generation. The model incorporates Chain-of-Thought (CoT) reasoning to translate high-level design requirements into structured microfluidic system descriptions.

The model generates outputs in a structured JSON format following a predefined schema (see: Output Format). The generated JSON describes a complete microfluidic chip, including:

• microfluidic components
• component parameters
• channel connections
• structural relationships between elements

This allows the model to act as a design file generator for microfluidic systems, enabling automated or AI-assisted microfluidic chip design workflows.

The repository includes:

• LoRA Adapter weights
• Quantized, split GGUF model files compatible with Ollama, may require merging before use

GGUF files can be merged using tools provided by llama.cpp (see: Merging Split GGUF Files).

---

Intended Use

This model is intended for:

• Automated microfluidic chip design generation
• AI-assisted CAD workflows for microfluidics
• Research in AI-assisted scientific design
• Programmatic generation of microfluidic device specifications

The model converts natural language design requirements into structured microfluidic design specifications.

Example Applications

• Rapid prototyping of microfluidic devices
• Automated generation of chip layouts
• Integration with microfluidic CAD pipelines
• AI-driven design exploration

---

Model Architecture

• Base Model: Qwen 3 32B
• Fine-tuning Method: Cold-start SFT LoRA
• Reasoning Strategy: Chain-of-Thought prompting and supervision
• Output Format: Structured JSON

The model is trained to produce schema-compliant structured outputs representing microfluidic chip configurations.

---

Output Format

The model generates JSON objects conforming to a predefined schema.

Schema definition:

https://github.com/TUM-EDA/Flui3d-Chat/blob/master/Dataset%20and%20Training%20Framework/datasets/resources/json_schemas/microfluidic_schema.json

The JSON output typically includes:

• Component definitions
• Channel connections
• Parameterized microfluidic elements
• Junction definitions

Example Output

{
  "connections": [
    {
      "source": "inlet_1",
      "target": "mixer_1"
    },
    {
      "source": "inlet_2",
      "target": "mixer_1"
    },
    {
      "source": "mixer_1",
      "target": "outlet_1"
    }
  ],
  "junctions": [
    {
      "id": "junction_1",
      "type": "T-junction",
      "source_1": "inlet_1",
      "source_2": "inlet_2",
      "target": "mixer_1"
    }
  ],
  "component_params": {
    "mixers": [
      {
        "id": "mixer_1",
        "num_turnings": 4
      }
    ],
    "delays": [],
    "chambers": [],
    "filters": []
  }

---

Repository Contents

This repository includes:

1. LoRA Adapter

The LoRA adapter can be loaded on top of the base Qwen model for inference or further fine-tuning.

2. Quantized GGUF Models

Quantized GGUF format models compatible with:

• Ollama
• llama.cpp

Due to file size limitations, the GGUF models are split into multiple parts. These files must be merged before use.

---

Merging Split GGUF Files

To merge the split GGUF files, use the merging utilities from llama.cpp:

https://github.com/ggml-org/llama.cpp/blob/master/tools/gguf-split/README.md

---

Usage with Ollama

The merged GGUF file can be used with:

• Ollama

Example prompt:

Design a microfluidic chip with two inlets, one mixer, and a single outlet.

---

Limitations

• The model assumes valid schema-based output format and may produce invalid JSON if prompts are poorly structured.
• Generated designs should be validated before fabrication.
• The model does not replace domain expert verification.

---

Citation

If you use this model in academic work, please cite:

WILL BE PUBLISHED