Shot-to-Shot Neural Control of Plasma States in Laser-Driven Experiments

High-repetition-rate (HRR) laser-driven experiments are transforming high energy density physics by enabling rapid, data-rich exploration of complex plasma phenomena. Real-time control in these systems remains difficult because iterative optimization algorithms and linear feedback loops (e.g., PID) consume shots during convergence and adapt poorly to nonlinear, drifting operating conditions. Here we demonstrate shot-to-shot direct ML control for laser-driven plasmas, implemented with a neural setpoint controller that maps a requested plasma state and measured operating conditions directly to actuator settings. Applied to blast wave position control on a multi-joule platform, the method delivers sub-second inference compatible with 1 Hz operation and maintains the target po- sition with sub-millimeter precision while background pressure varies by more than 400% within the trained pressure bounds. Benchmarking against a hybrid neural-network plus Bayesian-optimization framework shows a substantial gain in response speed and control authority. These results establish a practical route toward retrainable autonomous control for next-generation laser-plasma platforms.