FASTER
Autonomous experimentation at synchrotron and neutron facilities aims to accelerate materials discovery and deepen the understanding of complex phenomena by closing the loop between measurement and decision-making, enabling intelligent, real-time exploration of experimental parameter spaces. This requires robust, adaptive algorithms that can interpret incoming data immediately, quantify uncertainty, and select the next experimental settings to maximize information gain under practical constraints. Classical analysis methods—such as fast Fourier transforms and peak fitting—remain essential for reliable, low-latency data reduction and parameter estimation. However, machine-learning approaches, in particular Bayesian optimization and reinforcement learning, are key enablers of autonomous decision-making in complex, high-dimensional landscapes and for leveraging prior data to improve experimental efficiency. To satisfy stringent feedback-latency requirements, computationally intensive stages in the autonomous pipeline will be accelerated on FPGAs, especially where conventional data analysis or learning algorithms are too slow for real-time closed-loop operation.
Projectdescription
FASTER aims to make synchrotron and neutron experiments substantially more efficient by enabling closed-loop, autonomous measurements with fast feedback: experimental parameters are adjusted on-the-fly based on real-time data treatment, analysis, and decision-making. Starting from pilot deployments at PETRA III beamlines P08 and P10 (DESY) and REFSANS (MLZ/FRM II), the project targets three representative use cases—X-ray Photon Correlation Spectroscopy (XPCS), X-ray and neutron reflectivity, and time-resolved reciprocal-space mapping (TR-RSM)—and develops transferable software/hardware building blocks applicable across techniques. The core outcome is an autonomous data pipeline that performs background subtraction, denoising, masking/ROI selection, beam-damage detection, and parameter estimation, coupled to optimization algorithms (classical and ML-based, including Bayesian/GP and reinforcement-learning style decision makers) that steer the next measurement using use-case-specific cost functions and quality metrics. To meet sub-second (and ultimately near-real-time) feedback requirements at modern detector data rates, computational bottlenecks will be addressed via FPGA-based acceleration for demanding operations such as correlation-function computation and low-latency inference, integrated through standardized interfaces into facility control stacks (e.g., BLISS/Sardana and TANGO/NICOS) to support routine beamline operation. In parallel, curated experimental and simulated training datasets will be produced in community formats (e.g., NeXus) and made accessible via public repositories and dissemination channels, complemented by outreach, “common user mode” workflows for non-experts, and an industry-oriented data interface aligned with FAIR principles—together laying the groundwork for robust, deployable autonomous experimentation at large-scale research infrastructures.
Focus points/areas
-
HW Acceleration
-
Autonomus Experimentation
-
Optimization Algorithms
-
Machine Learning
.
Methodology
Define system requirements and interfaces (WP2→WP3): Derive latency/throughput/memory targets from the use-case metrics and ML workflow, and specify the detector-to-compute and compute-to-beamline feedback interfaces (data formats, synchronization, safety bounds).
Implement the embedded autonomous pipeline on FPGA (WP3): Build the end-to-end low-latency chain: high-rate data ingest (detector/DAQ → FPGA), real-time pre-processing (denoise, background subtraction, masking/ROI update), and a decision block that converts quality/damage/SNR/correlation indicators into self-adjusting experimental parameters, followed by deterministic control output drivers to the beamline.
Optimize and verify iteratively (WP3): Profile bottlenecks, optimize compute kernels and data movement, validate against offline reference results, and ensure real-time closed-loop deadlines are met under realistic detector rates.
Integrate and deploy (WP3→WP4): Perform initial end-to-end testing in the USIE lab (XPCS pipeline), then tailor configurations to beamline use cases (XPCS P10, TR-RSM P08, XRR/NR P08/REFSANS) and support final hardware integration into the project stack.
