Pit Density Forecasting for Abrasive Finishing Processes of Nuclear Fusion Target Capsules

Abstract

Inertial Confinement Fusion (ICF) experiments at Lawrence Livermore National Laboratory require target shells with nanometer-scale surface precision for successful ignition. These shells undergo ultra-precision lapping and polishing; however, the process is susceptible to rare, deep pits that can induce hydrodynamic instabilities and degrade fusion performance. Because shell fabrication is time- and resource-intensive, reliable early prediction of shell quality is critical for reducing costs and improving yield. However, predicting pit evolution is challenging due to the nonlinear, stochastic, and stage-dependent dynamics of the process, while existing approaches rely on scalar metrics that fail to capture rare defects and distributional changes. To address this limitation, we propose a distribution forecasting model that predicts the full pit distribution across polishing stages. The approach employs a dual-encoder CNN–Transformer architecture that integrates kernel density estimation (KDE)-based pit distributions with process parameters through cross-attention. The model is trained using a composite loss combining Kullback–Leibler divergence, Wasserstein distance, and L2 error to capture tail behavior and distribution shifts. Across 21 experimental batches, the method reduces KL divergence and Wasserstein distance by 27–66% compared to multiple baselines and achieves 88% sensitivity in detecting large, deep pits at a 22% false positive rate. The framework enables early rejection of poor quality shells and facilitates process plan optimization, advancing a transition from reactive inspection to proactive, data-driven manufacturing.