Multi-Resolution COokiebox (MRCO) Detector Denoising, Classification, and Regression Data and Models

This dataset supports research in real-time characterization of attosecond X-ray pulses, especially in single-shot diagnostics using angular streaking and specifically for the Multi-Resolution COokiebox (MRCO) detector at SLAC's LCLS-II. It includes (1) raw simulated diagnostic images, (2) preprocessed data and normalization scalers, and (3) trained machine learning model weights.

The raw images were generated using Cookie SimSlim, a simulator that produces 2D angle-energy distributions mimicking what is observed in experimental detectors. These simulations span a wide variety of physical scenarios, including varying numbers of X-ray sub-pulses and phase separations, incorporating realistic noise and jitter conditions.

The preprocessed data includes normalized versions of these images, with consistent spatial dimensions and saved in .npz format. Ground truth labels (e.g., number of sub-pulses, relative phase) are included, along with the fitted MinMaxScaler objects used for normalization. The dataset is split into training/validation and test sets for convenient model development.

Trained PyTorch model weights are also provided for three tasks: (1) a convolutional autoencoder for denoising raw traces, (2) a BiLSTM classifier that predicts the number of sub-pulses, and (3) ResNet-based regression models that estimate phase and phase separation between sub-pulses. These models are intended for use in high-throughput ultrafast X-ray diagnostics and serve as baselines for future development.

This dataset is shared under an academic research license to support reproducibility, benchmarking, and further model development in the field of machine learning for FEL-based ultrafast science.