A neural network-based synthetic diagnostic of laser-accelerated proton energy spectra

Abstract Machine learning can revolutionize the development of laser-plasma accelerators by enabling real-time optimization, predictive modeling and experimental automation. Given the broad range of laser and plasma parameters and shot-to-shot variability in laser-driven ion acceleration at present, continuous monitoring with real-time, non-disruptive ion diagnostics is crucial for consistent operation. Machine learning provides effective solutions for this challenge. We present a synthetic diagnostic method using deep neural networks to predict the energy spectrum of laser-accelerated protons. This model combines variational autoencoders for dimensionality reduction with feed-forward networks for predictions based on secondary diagnostics of the laser-plasma interactions. Trained on data from fewer than 700 laser-plasma interactions, the model achieves an error level of 13.5%, and improves with more data. This non-destructive diagnostic enables high-repetition laser operations with the approach extendable to a fully surrogate model for predicting realistic ion beam properties, unlocking potential for diverse applications of these promising sources.