AutoMat 是用于扫描传输的端到端代理框架 电子显微镜图像到晶体结构重建并 材料-财产分析。

AutoMat已被接受为ICML 2026.

当前发布中还包括一个明确的agent_qwen系系层。 标准化的绳索 自然语言请求录入 Markdown技能 和用于交互测试的小浏览器演示。

□ 资源

• 纸张:自动Mat:通过代理工具使用从显微镜中启动自动晶体结构重建
• 数据集: Hugging Face上的STEM2Mat
• 代码:yyt-2378/Automat

□ 核心管道

AutoMat 围绕四个科学模块组织:

模块 角色 | – – – – – – –

• MOE-DIVAESR * 模式适应性 STEM 解析和超解析 * QQ Template Retriever 物理限制结构-模板搜索 QQ | STEM2CIF | 原子坐标提取和CIF生成 |
• MatterSim Wrapper* 可选的放松和财产评价 *

毒剂的利用使这些步骤明确和可复制,而不是 将步骤选择完全留给 LLM 。

∮什么是新的∮

• agent_qwen 利用线可复制执行.

AutoMat

AutoMat is an end-to-end agent framework for going from scanning transmission electron microscopy (STEM) images to crystal-structure reconstruction and materials-property analysis.

AutoMat has been accepted to ICML 2026.

The current release also includes an explicit agent_qwen harness layer with standardized harness lines, natural-language request intake, Markdown skills, and a small browser demo for interactive testing.

Resources

• Paper: AutoMat: Enabling Automated Crystal Structure Reconstruction from Microscopy via Agentic Tool Use
• Dataset: STEM2Mat on Hugging Face
• Code: yyt-2378/AutoMat

Core Pipeline

AutoMat is organized around four scientific modules:

The agent harness makes these steps explicit and reproducible instead of leaving step selection entirely to an LLM.

What Is New

• agent_qwen harness lines for reproducible execution.
• Natural-language intake that parses user requests into structured harness inputs.
• Markdown skills under agent_qwen/skills/*/SKILL.md.
• Uncertainty and confidence estimation for reconstructed structures.
• CLI, RPC, and browser-demo entry points.
• Compatibility wrappers for the historical structure_recongnition import name.

Repository Layout

AutoMat-main/
├── agent_qwen/                         # Harness, runtime tools, skills, demo
│   ├── harness.py                      # Standard harness-line runner
│   ├── request_intake.py               # Natural-language request parser
│   ├── runtime_skills.py               # Python wrappers around pipeline tools
│   ├── confidence_estimation.py        # Uncertainty/confidence report logic
│   └── skills/*/SKILL.md               # Normative Markdown skills
├── agent_qwen_vl_rpc.py                # RPC entry point for harness/LLM modes
├── src/
│   ├── pipline_framework.py            # Original STEM -> CIF -> property tools
│   ├── preprocess_model/               # MOE-DIVAESR model code
│   └── structure_paired_reconstruction # Template and atom reconstruction code
├── src/structure_recongnition/         # Compatibility aliases
├── utils/fft_convert.py                # FFT helper for confidence estimation
├── tests/test_agent_qwen_harness.py    # Harness/intake tests
└── harness_runs/                       # Local outputs, ignored by git

Installation

A conda environment is recommended.