Physics-driven self-supervised optimization of freeform surface parameters

The optimization of freeform optical systems typically relies on designer experience or large training datasets, making it difficult to achieve stable design. In this work, we propose to overcome these limitations by introducing a differentiable physical forward model, which describes the optical propagation process, into a deep neural network. The resulting Physics-Driven Self-Supervised Optimization method (PDSO) requires no training data. Based on a given initial structure, it can perform local re-optimization of the parameters of prespecified freeform surfaces, while all other structural parameters remain fixed. This reduces manual intervention and reliance on design experience during the local re-optimization process. We validate this approach using three typical refractive and reflective optical systems, including Three-Mirror Off-Axis, Cooke Triplet, and Double-Gauss systems. The results show that, compared with a single DLS optimization in ZEMAX, PDSO achieves comparable or better results in terms of merit function value, RMS spot radius, full-field aberrations, and MTF for all three systems, while also demonstrating good convergence stability.