Simulation-informed residual correction learning method for deformation prediction and compensation in digital light processing three-dimensional printing

Dimensional deformation remains a major obstacle to accuracy control in digital light processing (DLP) printing, especially for freeform parts, for which real deformation labels require costly physical printing, scanning, registration, and deviation extraction. This study provides a systematic attempt to address deformation prediction and physical compensation of DLP printed freeform parts with limited real printed samples. A simulation-informed residual correction (SIRC) deep learning method is proposed, where simulated deformation labels are used to provide low-fidelity priors and real printed deformation labels are used to supervise the residual correction. The final deformation field is obtained by combining the simulation prior prediction and the predicted residual. To further improve residual correction, multilevel feature distillation transfers deformation features from the model trained on simulated deformation labels, while geometry-importance-guided residual weighting focuses supervision on geometrically sensitive regions. To support model training and validation, a multi-fidelity dataset of shoe last geometries was established, representing parts with complex freeform surfaces. The simulated samples were used to train the teacher network, while real samples were used to construct residual labels and train the student network. Physical inverse compensation validation showed that the proposed method reduced the average pointwise Euclidean error (AEE) from 0.4721 to 0.1691 mm, indicating that the method provides a physically validated solution for deformation prediction and compensation of DLP printed parts.