---

AutoMat

AutoMat focuses on characterization to property analysis.

✨ Project Overview

AutoMat is an end‑to‑end framework that enables academic researchers to go from scanning transmission electron microscopy (STEM) images to crystal‑structure reconstruction and finally to materials‑property prediction in one click. The core pipeline consists of four modules:

1. MOE‑DIVAESR — pattern‑adaptive STEM denoising & super‑resolution
2. Template Retriever — physics‑constrained structure‑template search
3. STEM2CIF — symmetry‑aware atomic‑coordinate / CIF generation
4. MatterSim Wrapper — machine‑learning potential for fast relaxation & property evaluation

For the accompanying paper and benchmark dataset, see STEM2Mat‑Bench (link in the paper).

---

🛠️ Quick Installation

A Conda environment is recommended. A ready‑to‑use img2struc.yaml is provided in the repository root:

# 1) Create an isolated environment (Python 3.10+ recommended)
conda env create -f img2struc.yaml
conda activate img2struc

# 2) (Optional) Install with requirements.txt
pip install -r requirement.txt

🔧 Training the MOE‑DIVAESR Model

You can either reuse the official STEM2Mat‑Bench on Hugging Face
(https://huggingface.co/datasets/yaotianvector/STEM2Mat) or prepare your own dataset following the folder convention below.

dataset/
└─ SRDATA/                 # root
   ├─ training/
   │   ├─ LR_original/     # low‑resolution STEM tiles
   │   └─ HR/              # matching high‑resolution (ground truth)
   ├─ validation/
   │   ├─ LR_original/
   │   └─ HR/
   └─ test/
       ├─ LR_original/
       └─ HR/

1. **Download / create** the images and place them in the corresponding
   `LR_original` and `HR` sub‑folders.

2. **Adjust config** in `configs/moe_model.yaml` (patch size, batch size,
   learning rate, etc.).

3. **Launch training**

   ```bash
   # from project root
   python src/ensemble_model_train.py \
          --config configs/moe_model.yaml \
          --data_root dataset/SRDATA

4. The script will save the best MOE‑DIVAESR weights under runs/models/. Point the agent to this checkpoint to enable custom denoising

---

🚀 Usage Example

AutoMat ships with an agent‑based entry script that lets you reconstruct a crystal from a STEM micrograph and predict its properties in a single command —or interactively via chat.

agent_based.py
└─ Improved Materials‑AI Agent
   ├─ Mode 1: one‑shot pipeline
   └─ Mode 2: interactive chat

### 1 · One‑shot pipeline

```bash
python agent_based.py \
  --api_key YOUR_OPENAI_KEY \
  --image_path /path/to/img.png \
  --work_root ./results \
  --user_message "Elements: Al, Sb; dose = 30k"

2 · Interactive chat

python agent_based.py --api_key YOUR_OPENAI_KEY

User> /run /path/to/img.png ./results Elements: Al, Sb
User> exit      # or quit

3 · Qwen-VL agent example（Recommend）

src/agent_qwen_vl.py offers a small chat agent that talks to a Qwen‑VL service via OpenAI style function calling. Each pipeline step (patch_infer, template_match, cif_from_label and relax_cif) is exposed as a tool and usage statistics are printed when you exit.

Specify the service endpoint and model name as below:

python src/agent_qwenvl.py \
  --api_key YOUR_TOKEN \
  --model qwen-plus-2025-04-28

Output folders

Folder	Contents
01_recon/	Denoised / super‑resolved STEM image
02_label/	Best‑matched structure template
03_recon_cif/	Reconstructed primitive unit cell (*.cif)
04_relaxed/	MatterSim‑relaxed structure + predicted properties (LLM-agent output, etc.)

All intermediate and final results are stored under --work_root.