A modular framework for collaborative human-AI, multi-modal and multi-beamline synchrotron experiments

TL;DR

This work tackles the inefficiency of traditional mapping in heterogenous, multi-modal materials by introducing a modular, open-source AI-assisted framework that coordinates real-time, multi-beamline measurements at NSLS-II via Bluesky. It couples AI agents for data reduction, analysis, and Bayesian-optimization-driven decision-making with human adjudicators to ensure safe, autonomous operation across beamlines. The authors demonstrate synchronous XRD and XAFS mapping on Al-Ni-Pt combinatorial libraries, supported by a digital twin that shows AI-driven strategies outperform conventional and random sampling in resolving phase boundaries and minority phases. The framework, built on open tools and designed for plug-and-play extensibility, lays the groundwork for routine AI-assisted, multi-modal experiments at large facilities and beyond.

Abstract

High-throughput materials discovery and studies of complex functional materials increasingly rely on multi-modal characterization performed at synchrotron light sources. However, measurements are typically done with no use of data until after an experiment, neglecting opportunities for data-driven insights to guide measurements. We developed a modular, open-source framework that incorporates artificial intelligence within the Bluesky control and data streaming infrastructure at NSLS-II, enabling real-time orchestration of multi-beamline, multi-modal experiments. AI agents perform on-the-fly reduction, clustering, Gaussian process modelling, and Bayesian optimization driven data acquisition, while users monitor agent behavior and visualize results live. Combinatorial libraries of the ternary Al-Ni-Pt system were spatially mapped by X-ray diffraction and X-ray absorption fine structure measurements at the PDF and BMM beamlines, respectively. Dynamic switching between AI-driven and conventional grid mapping strategies was achieved, demonstrating the flexible workflows possible through this framework. A digital twin constructed from a simulated Al-Li-Fe oxide dataset shows that AI-driven mapping strategies outperform conventional mapping as well as random sampling by prioritizing measurements that better resolve both phase boundaries and localized minority phases. This framework supports plug-and-play capabilities, and establishes a foundation for routine multi-modal, AI-assisted large-scale user-facility operations.

Figures (13)

• Figure 1: Complementary information content of X-ray diffraction patterns and X-ray absorption spectra. In a combined analysis the partial pair distribution functions from XAFS measurements provide chemistry-specific structural constraints such as bond-distances.
• Figure 2: Flow diagram of a conventional beamline architecture (outer loop) at NSLS-II and the incorporation of AI agents to the workflow. This generalized workflow is expanded upon in Figure \ref{['fig:multimodal_blockdiagram']} with details about agents used in autonomous experiments done at the PDF and BMM beamlines, as well as in silico.
• Figure 3: Bayesian optimization-based autonomous experiment workflow with examples of AI agents for data dimensionality reduction, model construction, and measurement suggestion. Tasks can be performed by many different AI agents (sublevels) individually or in concert, with agent outputs as input to subse-quent agents. This workflow is used for the decision-making layer (Figure \ref{['fig:multimodal_blockdiagram']}) that drives autonomous experiments at the PDF and BMM beamlines, as well as in silico, and is highly extensible.
• Figure 4: Al-Ni-Pt thin films prepared on an amorphous glass substrate mounted on the PDF beamline for XRD measurements (left) and BMM beamline for XAFS measurements (right).
• Figure 5: Block diagram of software used in multi-modal AI-driven experiments with human-in-the-loop to orchestrate synchronous XRD and XAFS measurements. All agent outputs are stored and visualized so human experts can evaluate decision-making as well as results from analysis including XRD and EXAFS modeling. This workflow is generalized in Figure \ref{['fig:simple_meas_loop']}, and further examples of agents for the decision-making layer are provided in Figure \ref{['fig:BO_loop']}