Heterogeneous Molecular-Silicon Integration for Precision Neuromorphic Computing.

Abstract

While an expanding repertoire of materials exhibits neuromorphic functionality, translating these advances into scalable hardware integrated with silicon remains a formidable challenge. Monolithic integration, though ultimately desirable, is rarely tractable at early stages of material development. Progress is therefore often assessed through idealized simulations that overlook system-level constraints and non-idealities such as device variability, mixed-signal conversions, peripheral circuit noise, and energy overhead, yielding projections that can mislead. Here we present a heterogeneously integrated platform as a necessary stepping-stone toward scalable molecular-silicon hardware-a 64 × 64 molecular crossbar coupled with custom Si-circuitry in which every interface and signal pathway is explicitly defined and directly measured, yielding a realistic map of the design space for further scale-up. We demonstrate a one-step vector-matrix multiplication with > 12-bit precision, ∼73 dB signal-to-noise ratio, ∼80-ns write speeds, and high write accuracy across 10 000 programming cycles without correction loops, a substantial advance over the state of the art in dot-product engines. Using this platform, we execute workloads spanning signal and image processing, biomarker detection and tracking, and cryptographic operations. This work establishes the foundation on which subsequent scale-up strategies, including monolithic molecular-silicon integration, can be designed and assessed.