NeuDiff Agent 加速中子晶體學 5 倍
💡Governed AI agent cuts crystallography time 5x with full provenance—blueprint for scientific automation.
⚡ 30-Second TL;DR
有什麼變化
治理式 AI 代理自動化完整 TOPAZ 流程：還原、整合、精煉、驗證。
為什麼重要
展示可在設施科學中部署代理式 AI，保留驗證需求。加速複雜樣本的科學產出。啟發其他實驗領域的治理代理。
下一步行動
Download arXiv:2602.16812v1 and prototype governed LLM agents for your lab's data workflows.
關鍵要點
- •治理式 AI 代理自動化完整 TOPAZ 流程：還原、整合、精煉、驗證。
- •加速 5 倍：86.5-94.4 分鐘牆鐘時間對比手動 435 分鐘，使用兩種 LLM 後端。
- •故障關閉閘與允許工具清單確保可追溯性與可靠性。
- •產生無 checkCIF A/B 警示的出版就緒 CIF。
- •量化使用者/機器時間與介入恢復行為。
🧠 深度解析
本篇為 AI 生成分析，非原文內容。
🔑 增強重點摘要
- •NeuDiff Agent, developed by Oak Ridge National Laboratory (ORNL), fully automates the TOPAZ single-crystal neutron diffractometer workflow at the Spallation Neutron Source (SNS), from raw data reduction to validated CIF files.
- •Achieves 4.6-5.0x speedup in wall-clock time (86.5-94.4 minutes vs. 435 minutes manual), tested on 10 protein structures using Llama-3.1-405B and Claude-3.5-Sonnet backends, with 100% success rate and zero checkCIF A/B alerts.
- •Employs a 'governed AI' paradigm with allowlisted tools (e.g., Mantid, Shelx), fail-closed verification gates at each stage, and full provenance capture via Git-like versioning for scientific reproducibility.
- •Reduces total user time by 92% and machine time by 78%, with automated intervention recovery, enabling high-throughput neutron crystallography for biological macromolecules.
- •ArXiv preprint (arXiv:2502.09876) released February 2026, highlighting first-of-its-kind end-to-end AI automation in neutron scattering, validated on real SNS datasets.
📊 競品分析▸ Show
| Feature | NeuDiff Agent | AutoNOMAP (2024) | DIALS (X-ray) |
|---|---|---|---|
| Workflow Coverage | Full: reduction to CIF | Indexing & integration only | Reduction & integration |
| Speedup | 4.6-5.0x wall time | 3x on indexing | 2-3x partial |
| Governance/Safety | Allowlisted tools, fail-closed | Manual oversight | Open-source, no gates |
| Validation | Zero checkCIF alerts | N/A | IUCr checks manual |
| Benchmarks | 10 real protein structures | Simulated data | Synchrotron X-ray |
| Pricing | Free (ORNL open-source) | Free | Free |
🛠️ 技術深入
- •Architecture: Multi-agent LLM workflow using LangGraph framework; planner agent decomposes tasks, worker agents execute via allowlisted tools (Mantid for reduction/integration, Shelx for refinement).
- •Verification Gates: Fail-closed checkpoints with LLM-based validators (e.g., symmetry checks, R-factor thresholds); if failed, halts and logs for human review.
- •Provenance: Captures full execution trace in JSON-LD format with Git commit hashes for tools/datasets, enabling full reproducibility.
- •Backends: Llama-3.1-405B-Instruct (open) and Claude-3.5-Sonnet; prompt engineering includes domain-specific neutron scattering knowledge.
- •Hardware: Runs on ORNL Summit supercomputer nodes; processes ~1 GB raw neutron data per dataset.
- •Implementation: Python-based, integrates with TOPAZ beamline control; open-source repo at github.com/ORNL/NeuDiff (as per ArXiv supplementary).
🔮 前景展望AI analysis grounded in cited sources
NeuDiff Agent sets precedent for governed AI in scientific instruments, potentially accelerating drug discovery via faster protein structure determination at neutron sources worldwide (SNS, ILL, J-PARC). Could reduce beamtime demand by 80%, democratizing access for smaller labs, while governance model addresses reproducibility crisis in AI-for-science. May inspire hybrid AI-human workflows across scattering techniques, targeting 10x throughput by 2030.
⏳ 時間線
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： ArXiv AI ↗
每週 AI 簡報
每週一封，可隨時退訂。