Description
Laser-driven ion beams offer intrinsically ultrashort bunches and extreme instantaneous dose rates that are attractive for multidisciplinary applications spanning radiobiology, dosimetry, radiation resilience and imaging [1]. However, translating these sources into reliable ‘user beams’ remains challenging because key beam properties (e.g. spectral shape, cut-off energy, divergence and dose) are commonly inferred from diagnostics that interrupt the beam, are rate-limiting, or are otherwise incompatible with application delivery, while the source itself is highly sensitive to shot-to-shot fluctuations in laser and target conditions. Machine learning (ML) techniques can be used to directly address some of these challenges.
Building on our recent demonstration of a deep neural-network ‘synthetic diagnostic’ that predicts the full proton energy spectrum (with uncertainty) from non-disruptive secondary measurements and shot metadata—removing the need for direct spectral measurement on every shot [2] —we will present progress towards a more complete surrogate diagnostic suite. Using data from a dedicated three-week campaign on the ELIMAIA user beamline, we are training improved predictors and extending the approach to additional beam metrics such as divergence. The resulting framework aims to provide accurate, real-time characterisation without disrupting beam delivery, while identifying the laser/target parameters that must be stabilised to achieve reproducible source performance for end users.
[1] H. Daido, H et al., “Review of laser-driven ion sources and their applications” Reports on progress in physics, 75(5), 056401 (2012)
[2] C. J. McQueen et al., “A neural network-based synthetic diagnostic of laser-accelerated proton energy spectra” Communications Physics 8, 66 (2025)
