Data-driven surface figure analysis and process optimization for full-aperture continuous polishing based on processing data

Full-aperture continuous polishing (FCP) is widely used in the ultra-precision fabrication of large-aperture planar optical elements, but surface-figure control still relies heavily on operator experience because of the complex coupling among process parameters, machine states, and environmental conditions. This paper presents a data-driven framework for surface-figure prediction and sequential process-parameter optimization in FCP. First, multi-source machining data are acquired and processed to extract mechanism-guided features. An eXtreme gradient boosting (XGBoost) model is developed to predict the post-machining surface-figure error, achieving an R 2 of 0.8929 and a mean absolute error of 0.1273 λ (λ= 632.8 nm) on a separate validation set. Second, the multi-iteration polishing process is formulated as a Markov decision process, and deep deterministic policy gradient (DDPG) is introduced for sequential process-parameter adjustment. Finally, a separate set of machining records is used to evaluate the trained policy in a record-driven simulation environment. In the final evaluation, the policy requires fewer than 10 simulated interactions on average, compared with a historical mean of 16.25 machining iterations. These results demonstrate the feasibility of the proposed framework.