Optical model compression learning for galaxy morphology classification with diffractive deep neural networks

Galaxy morphology is key to understanding cosmic evolution. While large astronomical surveys such as the Sloan Digital Sky Survey produce vast amounts of data, energy-efficient processing remains a challenge. Diffractive deep neural networks (D²NNs) offer a promising optical computing solution with high throughput and low power consumption. However, deep D²NNs suffer from optical attenuation, while shallow networks have limited representation capability. To address this trade-off, we extend the concept of knowledge distillation with optical-domain considerations and develop an optical model compression learning (OMCL) framework for D²NNs. Numerical simulations on SDSS galaxy images show that the compressed model achieves 69.25% classification accuracy, representing a 27.5% improvement over conventional training, and exhibits partial recognition ability for unseen categories. Further analysis reveals that the compressed model develops smoother and sparser phase patterns, leading to improved robustness to noise and stronger tolerance to low-precision phase quantization (e.g., 3-bit). These results indicate that transferring optical-domain features from deeper networks can effectively mitigate the limitations of shallow D²NNs, providing a practical pathway toward fabrication-friendly and high-performance optical neural networks for large-scale astronomical data analysis.