Abstract
Optical system prediction reliability is limited by sparse high-dimensional sampling, where uncovered regions cause unpredictable failures difficult to detect with existing methods. This paper proposes cavity detection and targeted supplementation using kernel density estimation to locate sparse regions and add samples. Experiments on a ten-dimensional misalignment space show successful coverage of about 80% of missing regions; the prediction mean error drops to 18.6% and the RMS to 31.6% of the corresponding values of the initial model. The method requires no physical priors and provides an independent sampling quality assessment tool.