Autonomous X-ray Fluorescence Mapping of Chemically Heterogeneous Systems via a Correlative Feature Detection Framework

TL;DR

X-AutoMap tackles the bottleneck of serial XRF imaging by introducing a correlative feature-detection framework that operates across multiple elemental maps to identify regions of interest based on co-localization and exclusion patterns. The method integrates classical computer vision with rule-based logic and real-time beamline control (Bluesky), enabling closed-loop coarse-to-fine scanning that autonomously targets compositionally informative features. Applied to a chemically heterogeneous urban $PM_{2.5}$ sample, it reduced acquisition time from over 44 hours to ~10 hours while revealing diverse particle chemistries including fully mixed and partially overlapping elemental domains. The approach demonstrates robust, low-user-intervention imaging workflows at synchrotron beamlines and lays the groundwork for future ML-guided probabilistic sampling to further boost efficiency.

Abstract

We present X-AutoMap, a modular framework for autonomous X-ray fluorescence (XRF) mapping that enables chemically informed targeting of regions of interest through a correlative feature detection strategy. The system integrates classical computer vision and rule-based logic to identify features based on spatial relationships across multiple elemental maps, rather than relying solely on intensity or morphology. Tight integration with the Bluesky control infrastructure at the NSLS-II Hard X-ray Nanoprobe (HXN) beamline enables real-time, closed-loop scan orchestration. Applied to a chemically heterogeneous urban PM2.5 sample, X-AutoMap reduced high-resolution acquisition time from over 44 hours to approximately 10 hours by targeting compositionally significant features identified from coarse scans. High-resolution results revealed diverse particle types, including fully mixed, partially overlapping, and spatially distinct multi-element structures, demonstrating the ability of the framework to isolate chemically relevant features with minimal user intervention. The framework supports interactive and autonomous modes, operates within hardware constraints via grid-based scanning, and is robust across varying sample conditions. Future extensions will incorporate machine learning and probabilistic sampling to further improve detection sensitivity and scan efficiency. X-AutoMap is currently in active use at HXN and provides a flexible foundation for scalable, intelligent imaging workflows at synchrotron beamlines.

Figures (4)

• Figure 1: Multi-modal feature detection and ROI merging strategy used for autonomous targeting. Features are first detected independently from multiple input channels (e.g., elemental maps) using classical computer vision methods. Each detected feature is converted into a standardized bounding box based on the scan resolution. Bounding boxes from all channels are then spatially merged and clustered to identify regions of multimodal significance. The final set of high-priority scan targets (highlighted in red) reflects spatial overlap or correlation between features across different channels.
• Figure 2: Overview of the X-AutoMap autonomous scanning workflow. The system operates in a loop beginning with a coarse, low-resolution XRF scan of a mosaic region. Elemental features are detected in real time using OpenCV-based computer vision, and the resulting regions of interest (ROIs) are passed to a queue server for targeted high-resolution scanning. All scans and detected ROIs are saved as JSON files. A graphical user interface (GUI) then stitches the coarse and fine scans, overlaying the ROIs for visualization and verification. This closed-loop approach enables automated, real-time targeting of chemically relevant features in large-area scans.
• Figure 3: Example of autonomous region selection using X-AutoMap. Composite XRF map of a chemically heterogeneous PM$_{2.5}$ sample, with Fe (red), Ca (green), and Si (blue) elemental distributions. White dashed boxes indicate regions automatically selected for high-resolution scanning based on correlative feature detection. Yellow dashed boxes highlight representative features that were not selected by the autonomous workflow. This image illustrates both the effectiveness and current limitations of the correlative detection strategy in identifying chemically relevant features across large, complex scan areas.
• Figure 4: Examples of high-resolution XRF scans autonomously selected by X-AutoMap, illustrating different types of elemental correlations and mixing behaviors in PM$_{2.5}$ particles. Each column represents a distinct particle class based on spatial and chemical relationships among Fe (red), Ca (green), and Si (blue). Type 1 shows adjacent but largely unmixed elemental regions, indicating physical aggregation of chemically separate particles. Type 2 particles show strong co-localization of all three elements, suggesting chemically or physically mixed phases. Type 3 features exhibit partial overlap between two elements with the third spatially distinct. Scale bars indicate approximate particle dimensions in microns. These variations highlight the diversity of particle compositions and the effectiveness of the correlative feature detection strategy in targeting chemically informative structures.