Wavelength-encoded neuromorphic inference enabled by microcavity MoS2 photodetector arrays

Biological vision systems tightly couple spectral sensing with temporal integration to extract task-relevant information with minimal data movement. In contrast, conventional optoelectronic vision hardware typically separates photodetection from electronic computation, incurring substantial latency and energy costs. Optical neural networks can alleviate this bottleneck, but many implementations are difficult to scale. Here, we propose a bio-inspired optoelectronic inference architecture based on a Fabry-Perot microcavity-integrated MoS2 photodetector array, in which sensing, weighting, and accumulation are unified within each pixel. Cavity-engineered wavelength selectivity encodes neural-network weights in the spectral domain, while the finite carrier lifetime of MoS2 enables analog temporal accumulation without external memory. The system achieves test accuracies of 99.6%, 94.8% and 94.0% on MNIST, CIFAR-10 and the Free Spoken Digit Dataset, respectively. Post-training optical Hessian pruning further reduces optical complexity while maintaining robustness. This architecture provides a compact route toward wavelength-aware in-sensor neuromorphic inference. Conventional vision hardware separates photodetection from computation, incurring substantial latency and energy costs. Chen and Di et al. report an optoelectronic architecture based on a microcavity-integrated MoS2 photodetector array that unifies sensing, weighting and temporal accumulation within each pixel.