Abstract
Diffractive neural networks offer high-speed and energy-efficient computation for artificial intelligence hardware, yet their application to time-series signal processing remains underexplored. Here, we propose an electromagnetic diffractive neural network framework for music style and sentiment classification. Time-series audio signals are converted into log-Mel spectrograms and processed using a multilayer metasurface-based DNN. Experiments on the GTZAN dataset achieved classification accuracies of 90.3% for five genres and 96.33% for three genres. Furthermore, sentiment classification on the POP909 dataset achieved 90.56% accuracy. These results demonstrate that diffractive neural networks can effectively process time-series data and provide a promising hardware-efficient solution for intelligent signal processing in communication and sensing applications.