AI-accelerated high-fidelity volume fabrication of reflow-compatible micro-optics via UV imprinting

Wafer-level optics (WLO) fabricated via UV imprinting enable scalable, cost-effective production of reflow-compatible micro-optics for compact optical modules in consumer electronics. However, form error (FE) introduced during the manufacturing process—primarily due to polymer shrinkage and mold deformation—necessitates 2 to 3 human-guided compensation cycles, significantly increasing development time and cost. This work presents an AI (artificial intelligence)- accelerated, physics-assisted prediction framework that reduces the development cycles by 30% while maintaining high replication fidelity. We propose a two-step cascaded prediction approach that explicitly decomposes the median FE into a dominant best-fit sphere (BFS) component and a higher-order rotationally symmetric irregularity (RSI), consistent with optical surface standards and physical deformation mechanisms in UV imprinting. In the first step, an AI model predicts the global spherical error as a low-dimensional scalar quantity from single-point diamond turning (SPDT) design parameters. In the second step, this physics-informed prediction is incorporated as an additional feature to regress the full FE profile, enabling focused learning of residual local deformations. The framework is evaluated on 69 distinct micro-optic designs using five-fold cross-validation and benchmarked against conventional single-stage AI models and human-guided compensation cycles. Results demonstrate that the proposed two-step approach achieves up to 3.3× improvement in peak-to-valley accuracy over single-stage prediction and closely approaches the accuracy of second-cycle human correction. Feature-importance analysis further confirms that the cascaded structure successfully decouples global shrinkage-driven deformation from local stress-induced irregularities, providing both improved prediction performance and physical interpretability. These results show that physics-assisted AI can eliminate one full development cycle for challenging designs, offering a practical pathway toward faster, lower-cost, and high-fidelity volume manufacturing of reflow-compatible micro-optics.