A Hybrid Neural Architecture: Online Attosecond X-ray Characterization

TL;DR

The paper addresses the need for real-time, single-shot attosecond-scale characterization of SASE sub-spikes in MRCO sinograms at high repetition rates. It introduces DCIPHR, a modular hybrid architecture that fuses CNNs, a BiLSTM-based sub-spike classifier, and ResNet-based phase regressors, trained on CookieSim synthetic data to denoise, count sub-spikes, and estimate phase delays. Key results show the denoiser achieving ~${SSIM-G}$ gains of about 29%, sub-spike counting accuracy above 96%, and a double-sub-spike phase difference RMSE of 0.04652 rad, translating to ~29.6 attoseconds at a 1.2 µm streaking wavelength; end-to-end inference latency is 168.3 µs with >10 kHz throughput on Groq hardware and potential scaling to 14 kHz with FPGA. This enables real-time streaming event selection and autonomous experimental control at MHz-class rates, supporting plans for XLEAP and digital-twin approaches in next-generation high-rate XFEL facilities like LCLS-II.

Abstract

The emergence of high-repetition-rate x-ray free-electron lasers, such as SLAC's LCLS-II, serve as our canonical example for autonomous controls that necessitate high-throughput diagnostics paired with streaming computational pipelines capable of single-shot analysis with extremely low latency. We present the Deterministic Characterization with an Integrated Parallelizable Hybrid Resolver architecture, a hybrid machine learning framework designed for fast, accurate analysis of XFEL diagnostics using angular streaking-based sinogram images. This architecture integrates convolutional neural networks and bidirectional long short-term memory models to denoise input, identify x-ray sub-spike features, and extract sub-spike relative delays with sub-30 attosecond temporal resolution. Deployed on low-latency hardware, it achieves over 10~kHz throughput with \SI{168.3}{\micro\second} inference latency, indicating scalability to 14~kHz with FPGA integration. By transforming regression tasks into classification problems and leveraging optimized error encoding, we achieve high precision with low-latency performance that is critical for real-time streaming event selection and experimental control feedback signals. This represents a key development in real-time control pipelines for next-generation autonomous science, generally, and high repetition-rate x-ray experiments in particular.

Figures (5)

• Figure 1: X-ray Detection and Aggregated Readout Circularly polarized IR streaking field creates a momentum shift in the x-ray photonionized electrons at the interaction point in the MRCO detector. At the front of the pulse train is a 2 SASE sub-spike x-ray shot, resulting in two ionization events that produce a sinogram with 2 sinusoids separated vertically, an effective phase shape which can be mapped to temporal separation of the sub-spikes based on streaking laser wavelength.
• Figure 2: DCIPHR Architecture a) noisy input sinogram images are fed into b) the autoencoder-based denoiser; these denoised sinograms are then concurrently streamed to the c) SASE sub-spike classifier network, the e) double sub-spike regression network, and the f) single sub-spike regression network where d) shows the details of the LSTM block inside the classifier network; the output from the classifier is used as the select input for the g) output selector which routes both the output from the classifier as well as the properly selected double sub-spike or single sub-spike output for downstream analysis.
• Figure 3: Denoiser Autoencoder Model Results a) and b) show error distribution histograms for the SSIM-G metric and MSE, respectively, and c) and d) show the model input and output for a double sub-spike SASE pulse with the respective SSIM-G and MSE values labeled.
• Figure 4: LSTM Network for Sinogram Featurization Classification a) cross entropy matrix for the test set and b) an example of true 2 sub-spike case predicted correctly as 2.
• Figure 5: Double Sub-Spike Phase Difference Regression Shows the true versus predicted phase values, the labels from the classifier for each data point, and the residuals for those correctly identified as 2 sub-spikes.