APEXA: agent-driven detector calibration and azimuthal integration of CeO2/LaB6 synchrotron diffraction data (APS 1-ID)

Pawan K. Tripathi; Hemant Sharma; Andrew Chuang; Mathew J. Cherukara

synchrotron interoperability benchmark software LLM experiment X-ray diffraction crystal structure machine learning high-throughput synchrotron, X-ray diffraction, detector calibration, HEDM, azimuthal integration, LLM agents, agentic AI oxides version control simulation synchrotron interoperability benchmark software LLM experiment X-ray diffraction crystal structure machine learning high-throughput synchrotron, X-ray diffraction, detector calibration, HEDM, azimuthal integration, LLM agents, agentic AI oxides version control simulation

Year

2026

Source Name

117dd613-cf08-4bf5-9fb0-8f080e859437

DOI

10.18126/tgg4-1m26 View on Datacite

Get the Data

Raw detector frames, analysis outputs, and evaluation traces supporting APEXA (Advanced Photon EXperiment Assistant), a multi-agent LLM framework that automates synchrotron data reduction from natural-language instructions. Contents

raw/ai_tune/calibration/ — CeO2 and LaB6 powder calibration series collected at APS Sector 1-ID (Varex area detector, 63 keV, sample-to-detector distance ~900 mm), spanning an attenuation and exposure sweep, with matched dark frames. HDF5 frames alongside per-run refined geometry, convergence histories, and diagnostic renderings.
figure_data/ — tabulated calibration geometry and adversarial-safety counts underlying the published figures.
results/day2_safety/ — adversarial motor-safety evaluation traces (JSON), covering four language models under both tool-enforced and prompt-only conditions.

Methodology Detector geometry was refined with MIDAS auto-calibration against NIST-traceable CeO2 and LaB6 lattice constants; the refined geometry then drove azimuthal integration to 1D patterns. All analyses were performed on previously acquired detector data; no instrument was operated as part of this work. The motor-safety traces were obtained against a simulated EPICS IOC, not physical hardware. Analysis code, the evaluation harness, and task definitions are in the APEXA repository (see Related URL).