Hardware-calibrated optical spiking neural network based on a programmable MZI mesh weight engine

Phase-change-material (PCM) synapses and all-optical spiking neurons provide a promising neuromorphic-computing framework for on-chip photonic neural networks. However, PCM weights mainly rely on absorption-based transmittance modulation, which is intrinsically more suitable for non-negative weights. When signed real-valued or complex-valued weight matrices obtained by backpropagation are implemented, differential synapse pairs are usually required, thereby increasing device count, insertion loss, and programming errors. This paper proposes a hybrid photonic-neural network simulation framework that employs a wavelength-division multiplexing (WDM) structure to implement optical signal fan-out and channel allocation. The weight matrices are implemented using a programmable Mach-Zehnder interferometer (MZI) array, and a hybrid offline-online configuration calibration strategy is employed to compensate for hardware errors. Matrix-level simulations on the MNIST dataset show that, under the proposed network architecture, the PCM differential-synapse model achieves 92.39% accuracy, whereas the uncalibrated MZI weight-matrix scheme reaches 87.15%. After online calibration, the MZI scheme recovers to 92.59%. Further matrix-fidelity analysis shows that the PCM baseline reaches approximately 96% average fidelity. The uncalibrated MZI scheme achieves about 78.00% fidelity, whereas calibration improves the fidelity of the two layers to 97.53% and 97.40%, respectively. These results demonstrate that the proposed method enhances the expressive power of the weight matrix and improves the correctability of hardware errors while retaining the advantages of the original optical pulse network architecture. This work provides a viable path toward the construction of scalable optical neural networks.