Fabrication-Aware Training for Two-Photon Polymerized Diffractive Deep Neural Networks

Diffractive deep neural networks (D2NNs) provide a promising pathway toward high-speed, energy-efficient optical computing. However, a substantial discrepancy between simulation and experiment remains when devices are fabricated by two-photon polymerization (TPP) for visible/near-infrared operation. This discrepancy is primarily due to fabrication constraints. In particular, the finite voxel size of the writing focus and the layer-by-layer slicing quantization attenuate high-spatial-frequency phase features and distort the designed wavefront. Here, a fabrication-aware training (FAT) strategy is introduced to incorporate these dominant constraints into the optimization loop through a differentiable model, thereby improving the experimental performance of two-photon polymerized D2NNs. In a handwritten digit recognition task, FAT increases the average experimental energy concentration ratio (ECR) from 6.15% to 11.50%. This reduces the simulation-to-experiment discrepancy by 79.03% while preserving the classification accuracy. These results demonstrate that fabrication-aware optimization substantially improves the functional fidelity of micro/nanoscale diffractive optical computing devices.