High-accuracy inverse optical design by combining machine learning and knowledge-depended optimization

With respect to knowledge-dependent approaches (KDAs) that require optimization in the high-dimensional parameter space, data-driven methods (DDMs) show remarkable generalization and diversity but commonly with unsatisfactory accuracy for complex systems. To overcome the imperfections of the KDAs and DDMs, we suggest a composite scheme by combining them, which not only alleviates the optimization burden but also presents a remarkable generalization and accuracy. This composite scheme as an example is applied to design one-dimensional photonic crystals (1DPCs) from the transmission spectra, which first determines the 1DPC type by a classification neural network, then predicts the layer thicknesses of that 1DPC by a generative adversarial network (GAN), and finally further optimizes the layer thicknesses by the KDA that is based on the method of least squares and starts from the results of the KDA. Numerical results yield that the third step can improve more than 12% for the prediction accuracy with respect to the GAN for complex 1DPCs, resulting in the overall successful prediction probability being able to reach 96.8%. Since the scheme combines the KDAs and DDMs, it has remarkable generalization and high accuracy and provides a potential alternative for the efficient inverse design.