NeuDiff Agent: A Governed AI Workflow for Single-Crystal Neutron Crystallography

Technical Deep Dive: NeuDiff Agent for Single-Crystal Neutron Crystallography

Overview

NeuDiff Agent is a governed AI workflow system for accelerating single-crystal neutron diffraction analysis from raw data to validated crystal structures. It combines an explicit workflow model, constrained tool execution, fail-closed verification gates, and a provenance-first audit trail to reduce instrument-to-publication latency while preserving facility governance requirements.

Problem & Context

• Neutron diffraction analysis workflows are well-established but complex, involving iterative decisions across configuration, reduction, integration, refinement, and validation stages
• Latency grows when intermediate artifacts, assumptions, and validation outcomes are hard to track, reproduce, and audit across tools and stages
• Facility governance requires that actions be constrained to approved tools, verification be fail-closed at workflow boundaries, and decisions/outcomes be auditable and replayable

Methodology

Workflow Model

• Formalizes the analysis workflow as an explicit state machine with typed variables spanning data access, reduction, integration, refinement, and validation
• Constrains tool invocation to an allowlist and enforces fail-closed verification gates at stage boundaries

Verification Gates

• Hard bounds on unit cell, wavelength, d-spacing, and resolution
• Consistency checks on UB matrices and orientation metadata
• Parsing of tool logs for failure signatures (non-convergence, unreasonable displacements)
• Workflow gating to block refinement and validation until prerequisites are met

Provenance Tracking

• Produces a per-run audit bundle capturing prompts, configuration, state, tool-call arguments/outputs, warnings, and timestamps
• Enables controlled replay for debugging and method verification

User Interface

• Browser-based front-end with tabular artifact rendering, dialogue interactions, and figure generation
• Interprets user intents, routes to appropriate actions, and requests explicit authorization for changes

Data & Experimental Setup

• Benchmark dataset: TOPAZ scolecite, representative of routine operation
• Fixed prompt protocol, controlled task definition, and repeated end-to-end runs with two LLM backends
• Baseline is a single manual user run, agent-assisted runs are 5 independent replicates per backend

Results

Productivity and User Effort

• Reduces end-to-end time from 435 min (manual) to 86.5 ± 4.7 min (Gemini 3.0 Pro) and 94.4 ± 3.5 min (GPT-OSS 120B)
• Partitions time into user time and machine time, documenting speedup from workflow mechanics rather than just model verbosity

Publication Readiness

• Treats validation as an explicit workflow stage, with fail-closed gates
• Resolves 6 level A, 0 level B, 5 level C, and 15 level G checkCIF alerts through 7 user-authorized interventions
• Final CIF has no level A or B alerts, meeting publication-facing criteria

Interpretation

• Acceleration is only defensible when paired with explicit governance and provenance
• NeuDiff Agent reduces avoidable rework by enforcing consistent execution, capturing provenance, and concentrating user attention on scientific judgments
• Enables facility deployment by aligning with requirements for constrained actions, fail-closed verification, and auditable decisions

Limitations & Uncertainties

• Benchmark focused on a single reference case - broader validation across varied datasets is an ongoing engineering activity
• Extending verification gates to handle more failure modes and instrument variations is an area for further research

What Comes Next

• Deployment and broader validation in the TOPAZ user facility
• Exploring methods to incorporate expert feedback and learning into the verification gate framework
• Investigating ways to scale the provenance-first, fail-closed approach to other neutron and X-ray crystallography workflows