Related Experiment Video
Updated: Aug 7, 2026
10:10
Neutron Crystallography Data Collection and Processing for Modelling Hydrogen Atoms in Protein Structures
Published on: December 1, 2020
NeuDiff Agent: a governed AI workflow for single-crystal neutron crystallography
Zhongcan Xiao1, Leyi Zhang1,2, Guannan Zhang3
1Neutron Scattering Division Oak Ridge National Laboratory Oak Ridge TN37831 USA.
Journal of Applied Crystallography
|August 6, 2026
Summary
NeuDiff Agent, an AI workflow for TOPAZ at the Spallation Neutron Source, significantly accelerates structural crystallography analysis. This AI tool reduces analysis time by over 4.5x, delivering validated crystal structures efficiently.
Area of Science:
- Crystallography and Materials Science
- Artificial Intelligence in Scientific Research
- Neutron Scattering and Structural Analysis
Background:
- Large-scale scientific facilities face analysis and reporting latency, hindering throughput in structural studies.
- Iterative reduction, integration, refinement, and validation steps are time-consuming in traditional crystallography workflows.
- The Spallation Neutron Source (SNS) requires efficient data processing for structural studies.
Purpose of the Study:
- To introduce NeuDiff Agent, a governed, AI-driven workflow for automated crystallographic analysis at SNS.
- To significantly reduce the time-to-result and improve analysis efficiency for structural studies.
- To generate validated crystal structures and publication-ready Crystallographic Information Files (CIFs).
Main Methods:
- NeuDiff Agent utilizes a large language model (LLM) to coordinate established crystallographic tools.
- The workflow incorporates explicit governance: allowlisted tools, fail-closed verification gates, and complete provenance capture.
- Performance was benchmarked using fixed prompts and repeated end-to-end runs with two LLM backends, measuring time and intervention burden.
Main Results:
- NeuDiff Agent reduced wall time from 435 minutes (manual) to 86.5–94.4 minutes (4.6-5.0x faster) in a reference-case benchmark.
- The AI-generated CIFs were validated with no 'checkCIF' level A or B alerts.
- User and machine time were partitioned, and intervention burden/recovery behaviors were quantified.
Conclusions:
- NeuDiff Agent demonstrates a practical and effective deployment of agentic AI in facility crystallography.
- The workflow successfully preserves traceability and publication-facing validation requirements.
- This AI approach offers a significant improvement in analysis efficiency and time-to-result for structural studies.
