Harnessing Machine Learning for Single-Shot Measurement of Free Electron Laser Pulse Power

TL;DR

A machine learning model is developed that predicts the temporal power profile of the electron bunch in the lasing-off regime using machine parameters that can be obtained when lasing is on and that showed superior predictions compared to the state-of-the-art batch calibrations.

Abstract

Electron beam accelerators are essential in many scientific and technological fields. Their operation relies heavily on the stability and precision of the electron beam. Traditional diagnostic techniques encounter difficulties in addressing the complex and dynamic nature of electron beams. Particularly in the context of free-electron lasers (FELs), it is fundamentally impossible to measure the lasing-on and lasingoff electron power profiles for a single electron bunch. This is a crucial hurdle in the exact reconstruction of the photon pulse profile. To overcome this hurdle, we developed a machine learning model that predicts the temporal power profile of the electron bunch in the lasing-off regime using machine parameters that can be obtained when lasing is on. The model was statistically validated and showed superior predictions compared to the state-of-the-art batch calibrations. The work we present here is a critical element for a virtual pulse reconstruction diagnostic (VPRD) tool designed to reconstruct the power profile of individual photon pulses without requiring repeated measurements in the lasing-off regime. This promises to significantly enhance the diagnostic capabilities in FELs at large.

Figures (4)

• Figure 1: Process workflow. During beamline operation, we obtain measured "machine parameters" and the longitudinal phase space ("phase image") for each photon pulse. We use a multi-layer perceptron machine learning model ("MLP model") to predict the temporal power profile of the electron bunch if it would be obtained without lasing ("lasing-off electron power"). From the phase image, we calculate the temporal power profile of the electron bunch with lasing ("lasing-on electron power"). Thus, we can estimate the temporal power profile for each individual photon pulse ("Photon power") by subtracting the measured lasing-on electron power from the predicted lasing-off electron power. The top row (machine parameters $\rightarrow$ MLP model $\rightarrow$ lasing-off electron power) is the part of the process that we solved in this paper.
• Figure 2: MLP model training performance. (a) Training and validation loss. The validation loss is lower than the training loss, because we use a dropout of 0.43 during training. (b) Predictions for individual shots (blue line) matched the actual measurements (red line) better than measurements from previous shots (dashed green line) and better than the mean of all measurements in the training data (dotted orange line). (c) Boxplots of all mean squared errors in the test dataset. (Prediction) Mean squared error between the predictions and the measurements in the test dataset. (Mean) Mean squared error between the measurements in the test dataset and the mean of all measurements in the training dataset. (Neighbors) Mean squared error between adjacent measurements in the test dataset. The numbers at the right of (c) represent the median (interquartile range) of the errors as well as the number of observations $n$.
• Figure 3: Schematic layout of the FLASH facility at Deutsches Elektronen-Synchrotron (DESY), Germany. Data sources for training the machine learning model are indicated by dashed arrows.
• Figure 4: De-jittering. (a) Phase image. (b) Temporal power profile created by weighing the phase image by the energy axis and projecting onto the time axis. (c) Temporal power profiles for 700 samples before de-jittering. (d) Temporal power profiles after de-jittering.