Deep learning-based misalignment error correction method for spherical surface testing

Accurate calibration of misalignment errors is essential for high-precision interferometric testing of large numerical aperture (NA) spherical surfaces, as misalignment-induced aberrations degrade wavefront measurement accuracy. To improve the accuracy and efficiency of misalignment calibration, this paper proposes a deep-learning-based wavefront misalignment correction method using interferometric fringe patterns as input. Simulated datasets covering single- and multiple-misalignment conditions were generated, and two ResNet-50-based networks were developed using a classification-guided decoupled regression strategy. Simulation results on the standard in-range test sets show that magnitude prediction errors remain within 6% for all evaluated cases, and the residual wavefront RMS is substantially reduced after network-only correction in the single-misalignment simulation test. Additional simulations with perturbations at NA=0.9 show that the proposed method retains effective correction capability within the investigated perturbation ranges. Physical experiments conducted at NA=0.42 using separately retrained networks provide preliminary physical validation of the proposed method under the available experimental conditions.