Research on Scalable Complex-Valued Optical Convolution Based on Four-Wave Mixing

To address the limitations of existing optical computing schemes in handling complex-valued data, kernel scalability, and energy efficiency, this study proposes an optical convolution method based on the four-wave mixing (FWM) effect. A high-speed simulation platform operating at $10 \mathrm{Gbit} / \mathrm{s}$ was developed to comprehensively evaluate its performance. Compared with traditional implementations that rely on optical frequency combs and are typically constrained to kernel sizes of fewer than 70 points, the proposed method successfully achieves a 128-point complex-valued autocorrelation computation. The results demonstrate high computational accuracy, with the magnitude of the complex correlation coefficient between the experimental and theoretical outcomes reaching 0.93. Moreover, the method achieves a low per-bit energy consumption of 0.631 pJ/bit, significantly outperforming comb-based solutions. Further analysis reveals two main factors limiting system performance: (1) when the frequency spacing between the input and reference optical signals is too narrow, spectral overlap leads to crosstalk, reducing the correlation coefficient to $\mathbf{0. 7 6 3}$; and (2) when the two carrier frequency offsets satisfy a harmonic relationship, interference between FWM-generated components occurs, further degrading the correlation coefficient to 0.455. The findings indicate that careful design of carrier frequency offsets is essential for mitigating interference and enabling low-power, scalable complex-valued optical convolution.