Polishing is a critical process in fabricating space-telescope mirrors because it determines the surface figure and consequently optical performance. Deterministic polishing relies on the tool influence function (TIF), which describes the spatial material-removal profile. At nanometric removal depths, the TIF becomes highly sensitive to process conditions, limiting the accuracy of analytic models such as Preston’s equation. In this study, we propose a deep learning-based approach to predict TIF depth for polishing a Silicon Carbide (SiC) mirror surface. To mitigate data scarcity, we augment 231 experimental measurements with Gaussian noise consistent with the repeatability observed in repeated trials (≈ 20 nm peak-to-peak). The resulting model achieves a validation mean absolute error (MAE) of 4.24 nm and a test MAE of 3.99 nm; on nine additional experimental cases, the MAE is 6.75 nm. These results indicate that the proposed augmentation improves robustness to experimental variability and supports the development of a data-driven, automated polishing workflow.
Han et al. (2026) studied this question.