Co-evolving AI Optimizes Inertial Confinement Fusion Yield

Optimizing Inertial Confinement Fusion (ICF) is critical for achieving sustainable energy, but traditional offline-trained surrogate models often fail when iterative optimizers push inputs into regions outside their training distribution. This leads to unreliable predictions and hinders progress. Researchers have introduced Co4ICF, a novel co-evolving framework designed to address this challenge. Co4ICF couples a physics-informed surrogate model with a PPO (Proximal Policy Optimization)-based reinforcement learning pulse optimizer. The surrogate model is continuously fine-tuned using trajectories generated by the optimizer, effectively correcting extrapolation errors as the optimizer explores new input distributions. This allows the optimizer to query an increasingly accurate and adaptive environment. In 1D MULTI environment simulations, Co4ICF achieved a remarkable 146.1% normalized yield compared to current laser design baselines. Even more significantly, when the optimized pulse was evaluated directly in a 2D-MULTI environment without any specific 2D training or fine-tuning, it attained an astounding 246.9% normalized yield. These results, supported by budget-matched ablations, confirm that the co-evolving mechanism is crucial for these substantial gains, offering a promising path for ICF research.