Integrated AI/ML and HPC Framework for Multidimensional Neutron Crystallography
Structural Dynamics · 01.03.2025 · 12 (2_Supplement) · A30-A30
DOI: 10.1063/4.0000339 ↗Kurz gefasst
Fachartikel in der Zeitschrift Structural Dynamics, online erschienen am 10.04.2025. Verfasst von Xiaoping Wang und 11 weiteren Autor:innen. Die Arbeit wurde bisher 2-mal zitiert (Crossref). Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Neutron time-of-flight (TOF) event data at the Spallation Neutron Source (SNS) provides a wealth of temporal information in multidimensional diffraction and parameter spaces. However, current analysis methods rely on sequential steps that can limit the potential for on-the-fly analysis of complex single-crystal neutron diffraction patterns and hinder the seamless integration of knowledge across different analysis techniques and workflows. To address this, we present an integrated framework for real-time single-crystal neutron diffraction data reduction and analysis. Our approach combines advanced AI/ML models with high-performance computing (HPC) at the Oak Ridge Leadership Computing Facility (OLCF). This integration enables seamless experiment steering and real-time decision-making, optimizing the use of valuable neutron beam time. Our primary focus is on TOPAZ, a high-resolution single-crystal TOF Laue diffractometer at SNS. We analyze live 4D spatiotemporal data (3D in Q-space with time as the fourth dimension) for precise, voxel-level prediction of neutron scattering patterns. The core of our method is a Temporal Fusion Transformer (TFT) model, a specialized deep neural network (DNN) tailored for multi-horizon forecasting tasks. This model, embedded within a Markovian stochastic framework, offers unprecedented accuracy and interpretability in predicting neutron scattering patterns at the voxel level across temporal 4D spaces. We showcase the feasibility of this approach using a hierarchical parallelization scheme on the OLCF Frontier supercomputer with a subset of neutron TOF event data from the TOPAZ beamline. Our results demonstrate the transformative potential of integrating AI/ML and HPC for real-time experiment steering and decision-making in multidimensional diffraction and parameter spaces. This work signifies a substantial leap in neutron scattering research, paving the way for greater automation and efficiency. Ultimately, this integrated method, aligning BES's neutron facilities with ASCR's HPC resources using AI/ML, is a crucial component of a future Integrated Research Infrastructure (IRI). We invite collaboration to explore opportunities further and redefine the future of neutron science together.
Zitieren
APA
Wang, X., Zhang, G., Yin, J., Liu, S., Reshniak, V., Morgan, Z., Pawledzio, S., Proffen, T., Cao, H., Chakoumakos, B., Hoffmann, C., & Liu, Y. (2025). Integrated AI/ML and HPC Framework for Multidimensional Neutron Crystallography. Structural Dynamics, 12(2_Supplement), A30-A30. https://doi.org/10.1063/4.0000339
Vancouver
Wang X, Zhang G, Yin J, Liu S, Reshniak V, Morgan Z, et al. Integrated AI/ML and HPC Framework for Multidimensional Neutron Crystallography. Structural Dynamics. 2025;12(2_Supplement):A30-A30. doi:10.1063/4.0000339
BibTeX
@article{Wang2025,
author = {Wang, Xiaoping and Zhang, Guannan and Yin, Junqi and Liu, Siyan and Reshniak, Viktor and Morgan, Zachary and Pawledzio, Sylwia and Proffen, Thomas and Cao, Huibo and Chakoumakos, Bryan and Hoffmann, Christina and Liu, Yaohua},
title = {Integrated AI/ML and HPC Framework for Multidimensional Neutron Crystallography},
journal = {Structural Dynamics},
year = {2025},
volume = {12},
number = {2_Supplement},
pages = {A30--A30},
doi = {10.1063/4.0000339}
}
EUVIMED