Inverse Modeling of Laser Pulse Shapes in Inertial Confinement Fusion with Auto-Regressive Models

Realizing practical fusion energy remains one of society’s most significant unresolved scientific challenges, carrying profound implications for sustainable, carbon-free power. A key determinant of success in Inertial Confinement Fusion (ICF) experiments is the design of a Laser Pulse (LP) Shape capable of optimally driving implosions within strict physical limits. Conventional LP design depends on costly simulations and labor-intensive iterative tuning. To address this, we introduce the Laser Pulse Shape Design System (LPDS), a generative inverse modeling framework based on auto-regression that directly maps desired fusion outcomes and target pellet parameters to optimized LPs. We explore a multi-objective training setup to design diverse LPs that adhere to physical constraint while achieving less than 2\% error in the desired implosion outcomes. In addition, we incorporate constraint-conditioning via inpainting and gradient-based editing strategies, enabling precise control over pulse characteristics during generation. This framework offers a data-driven solution for LP design in ICF, advancing the pursuit of practical, sustainable fusion energy.