28 June 2026 to 2 July 2026
Ghent, Belgium
Europe/Brussels timezone

Online Serial Crystallography Classification at 35,000 frames per second with FPGA-Deployed Embedding Neural Networks

1 Jul 2026, 14:30
20m
Oehoe (Coupure Blok E)

Oehoe

Coupure Blok E

Oral presentation Oral presentations

Speaker

Luca Scomparin (SLAC National Accelerator Laboratory - Stanford University)

Description

The LCLS‑II X‑ray free‑electron laser facility is set to deliver an unprecedented repetition rate of up to 1 MHz, enabling novel scientific and discovery capabilities. Fully harnessing these capabilities requires imaging detector systems and data acquisition (DAQ) to sample and process data at correspondingly high frame rates. The ePixUHR family of detectors targets continuous acquisition of megapixel images at 35 kHz, while the future SparkPix family is designed to achieve MHz rates [1]. In this context, on‑detector calibration opens a path for machine‑learning workflows to be executed directly at the detector level.

In serial crystallography experiments, protein crystals are suspended in a stream of supporting liquid that is continuously intercepted by the X‑ray beam [2]. When a crystal is hit, it produces a diffraction pattern of interest; however, fluctuations in X‑ray pulse power and diffraction from the liquid support make discrimination between hits and misses a complex classification task. Discarding images with no diffraction pattern would provide a sizable reduction in storage and computing requirements, helping pave the way toward future MHz‑rate systems. An ePixUHR camera consists of multiple modules, each composed of six ePixUHR ASICs arranged in a 3×2 pattern and an acquisition FPGA. Because the full camera image is assembled later in the DAQ system, classification models must be designed with this natural segmentation in mind.

In this work, we present the design and characterization of a complete serial‑crystallography edge‑ML online classification pipeline. We follow the approach described in [2], with special attention on developing preprocessing methods that can be efficiently deployed to an FPGA. On‑detector calibrated sub‑images from each module are normalized, downsampled, stitched and then fed into a CNN‑based model, which is deployed on the DAQ FPGAs using the SLAC Neural Network Library [3]. The neural network produces an embedding vector that is passed to the DAQ; embeddings from the detector modules are then combined to produce a hit/maybe/miss classification, following the approach in [2]. The system has been deployed to a Varium C1100 evaluation card, where it utilizes less than 30% of the available configurable logic blocks, and less than 5% of the digital signal processors. The firmware was then validated using data from [2] whose geometry was adapted to match the one of ePixUHR 4-megapixel. Matching with software was in the order of 32-bit floating point machine precision. Frame rates in excess of 43 kHz could be sustained. We will also describe validation measurements of the detector performed as part of this effort.

References
[1] H. Sandberg et al., 2025 JINST 20 P08019
[2] T.-W. Ke et al., (2018). J. Synchrotron Rad. 25, 655-670.
[3] R. Herbst et a., Springer, pp. 120–134, 2022

Acknowledgements
R&D at the Linac Coherent Light Source (LCLS), SLAC National Accelerator Laboratory, is supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences under Contract No. DE-AC02-76SF00515

Authors

Abhilasha Dave (SLAC National Lab) Angelo Dragone (SLAC National Accelerator Laboratory (US)) Dawood Alnajjar (SLAC National Accelerator Laboratory (US)) Dionisio Doering (Slac National Lab.) Emmenuel Essel (SLAC National Accelerator Laboratory - Stanford University) Guilherme Paulino (SLAC National Accelerator Laboratory - Stanford University) James John Russell (SLAC National Accelerator Laboratory (US)) Larry Lou Jr Ruckman (SLAC National Accelerator Laboratory (US)) Laura King (SLAC National Accelerator Laboratory - Stanford University) Luca Scomparin (SLAC National Accelerator Laboratory - Stanford University) Ryan Taylor Herbst (SLAC National Accelerator Laboratory (US))

Presentation materials