Machine Learning based Data Analysis Accelerator Directorate Brendan O’Shea

Topics Ghost Imaging Neural Networks Reduce data requirements Add measurement capabilities Neural Networks Speed up analysis Reduce real time computational requirements

Data Analysis: Ghost Imaging Ghost Imaging / Single Pixel Camera Philosophy: don’t average out fluctuations… use fluctuations to probe your sample!  “Compressive ghost imaging” leverages the ML toolbox. = B A x Images, A target reconstruction B = [ 5037, 4783, 4891, … ] Bucket measures total transmission target Random patterns Siqi Li

Data Analysis: Ghost Imaging Ghost imaging with electrons photocathode quantum efficiency UV laser DMD Two Differences: Now cathode itself is the target No DMD, instead exploit the natural cathode laser variation bucket detector Cathode laser Cathode target # of e- gun, solenoid target Reconstructed QE: Siqi Li

Data Analysis: Ghost Imaging Ghost imaging in the time domain Pump X-ray target A. Picon Delay (fsec) Electron energy (eV) Target (x) Array of buckets (B) Sample, CO ProbeX-ray e- Time As system (CO) evolves after absorbing X-ray, electron energies change Incident radiation (A) Daniel Ratner

Topics Ghost Imaging Neural Networks Reduce data requirements Add measurement capabilities Neural Networks Speed up analysis Reduce real time computational requirements

Data Analysis: Computer vision for XFEL power Power Prediction from with Vision-based Neural Network XTCAV provides best record of X-ray power profile Before lasing Problem: Can only measure one image at a time e- energy (g) Alternative: Use computer vision on only lasing-ON image! After lasing e- energy (g) Power (W) Time (fs) Xinyu Ren

Data Analysis: Computer vision for XFEL power Power Prediction from with Vision-based Neural Network XTCAV provides best record of X-ray power profile Problem: Can only measure one image at a time Alternative: Use computer vision on only lasing-ON image! Unlabeled data Labeled Training Set Convolutional neural network (CNN) recovers X-ray power with smaller errors than current state-of-art Training Set Genesis Simulation CNN Experiment GAN Feed Training Data Data Augmentation The generative adversarial network (GAN) generates labeled images  Augment training for CNN or other data-hungry algorithms Xinyu Ren

Data Analysis: Temporal electric field reconstruction Input power Power reconstruction Input spectral power Maxwell, Timothy J., et al. International Society for Optics and Photonics, 2014. Phase reconstruction a) Goal: full reconstruction of a femtosecond X-ray pulse (both power and phase) b) Problem: temporal power diagnostics have bad resolution. No way to measure precisely.  No way to retrieve phase c) Proposed solution: combine temporal+spectral diagnostics with iterative optimization... or with a neural network! Xiao Zhang

Data Analysis: Prediction on simulated SASE dataset Datasets: Comparison with traditional iterative reconstruction: Lasing off beam data Simulated power profiles Genesis Power: Phase: NN 0.08% Iterative 0.22% Power error NN score: 0.83 Iterative score: 0.80 Phase score Data Preprocessing: normalization of input temporal power, spectral input power and the corresponding groundtruth. Then add smearing and noise Networks: Metrics: Power Phase NN reconstruction: ~0.2ms per prediction Iterative reconstruction: ~30s per prediction Power Phase Xiao Zhang  NN: ~ 105 faster

Edge Radiation Diagnostic Filter Quantify emittance and energy spread in a non-destructive manner Edge Radiation based diagnostics naturally integrates into present linac designs Diagnostic segments accelerator for better global tuning Integrates into Gaussian Process, optimizer and other methods Boost diagnostic speed and capabilities using Neural Network A single shot, non-intercepting diagnostic ideal for efficient computer control of accelerators Brendan O’Shea