Deep Learning for Solving Optical Inverse Problems: CNN-Based Wavelength Prediction and Phase Retrieval Techniques

This project was released by sanarhm on GitHub on June 15, 2026. Its core is to use deep convolutional neural networks (CNN) to solve optical inverse problems, enabling end-to-end prediction of incident light wavelengths from Fraunhofer diffraction patterns. The project uses the PyTorch framework and aims to break through the limitations of traditional methods. Original link: https://github.com/sanarhm/Deep-Learning-for-Optical-Pattern-Recognition-and-Phase-Retrieval-in-Engineering-Physics

Optical inverse problems are classic challenges in engineering physics. Inverting wavelengths/phases from Fraunhofer diffraction patterns is an ill-posed problem (multiple solutions). Traditional methods like the Gerchberg-Saxton algorithm have high iteration costs, are prone to local optima, and are sensitive to noise. Deep learning provides a new direction for solving such problems through data-driven end-to-end learning.

Based on Fraunhofer diffraction theory (far-field diffraction, where wavelength is inversely proportional to the characteristic scale of the pattern), the project uses PyTorch to build a CNN: input preprocessed diffraction intensity patterns, extract spatial features (edges, textures, frequency distributions) via convolutional layers, output wavelength predictions (regression task), bypassing the complexity of traditional iterations.

Due to the difficulty in obtaining real data, synthetic data (labeled diffraction patterns generated based on physical principles) is used, covering various conditions (wavelength ranges, aperture shapes, noise). Data augmentation includes adding Gaussian noise, random rotation and scaling, adjusting contrast and brightness to improve model robustness.

For the regression task, MSE/MAE loss functions are used; the optimizer is Adam (adaptive learning rate); learning rate decay and early stopping mechanisms (monitoring overfitting with the validation set) are adopted, and the dataset is divided into training/validation/test sets.

Compared to traditional iterative algorithms, the trained model can predict with a single forward pass (milliseconds vs. seconds/minutes); no explicit physical modeling is needed, reducing reliance on domain knowledge; trained with noisy data, it has stronger noise resistance than traditional methods.

Application scenarios: Real-time spectral analysis (industrial monitoring), phase retrieval imaging (X-ray crystallography, etc.), optical system calibration. Limitations: Generalization ability (accuracy decreases beyond training range), interpretability (black box), insufficient multi-parameter joint estimation. Future directions: Improve extrapolation ability, introduce explainable AI, support multi-task learning.

This project demonstrates the potential of deep learning in solving optical inverse problems. CNN end-to-end learning enables wavelength prediction, with significant advantages in efficiency and noise resistance, providing a new research direction for optical inverse problems and promising applications in more engineering physics scenarios.