Towards AI-driven Experiments at PW-class Laser Facilities

Today, there are multiple high-intensity short-pulse lasers around the world that are capable of operating at high repetition rate (>1 Hz), representing an opportunity to accelerate the rate of scientific exploration by >3 orders of magnitude. In order to achieve this, diagnostics, targeting, laser control, and diagnostic analysis must all operate at commensurate rates. Machine learning provides a path for achieving this by utilizing fast surrogate models for each piece. While demonstrations of this technology for these purposes have been growing in number, they must now be integrated into a fully autonomous system through artificial intelligence.

Here, we present progress on this front by utilizing a physics-based ML model as the guide for experimental exploration and optimization by; proposing experimental samples, analyzing data, retraining the original model, and proposing new experimental samples with a “human in the loop”. The experiments were carried out at CSU’s ALEPH laser facility where the laser is controlled through spectral phase shaping and MeV-energy electrons and protons are measured with HRR diagnostics. The process was iterated multiple times in order to reduce the model uncertainty while searching for optimal laser settings to increase MeV particle production.