Description
Machine‑learning techniques are emerging as powerful tools for advancing laser‑plasma accelerators, enabling real‑time optimisation, predictive modelling and increased automation of experimental campaigns. In high‑repetition‑rate laser‑driven ion acceleration, the large parameter space and significant shot‑to‑shot variability create a critical need for continuous, non‑disruptive diagnostics capable of providing real‑time feedback on ion beam properties. To address this challenge, we present a neural‑network‑based synthetic diagnostic designed to predict the energy spectrum of laser‑accelerated protons using only measured laser pulse parameters and secondary observables of the laser–plasma interaction. The approach integrates a variational autoencoder for efficient dimensionality reduction with a feed‑forward neural network that maps reduced diagnostic features to proton spectral outputs. Trained on fewer than 700 experimental interactions, the model achieves a prediction error of 13.5%, with accuracy improving systematically as additional data are incorporated. This non‑destructive diagnostic is fully compatible with high‑repetition‑rate operation and provides a foundation for surrogate models capable of accurately predicting ion‑beam properties in real time, significantly enhancing the stability and applicability of laser‑driven ion sources.
