Heterogeneous 3D Integration Based on Atomic-Level Electronics

diffusive memristors produce threshold-driven, biologically inspired spiking. These layers form a compact neuromorphic stack capable of synaptic plasticity, firing-rate modulation, temporal learning, analog in-memory computation, and convolutional feature extraction. This architecture allows layers to be manipulated enabling on-demand compute scaling. Demonstrations of analog vector-matrix multiplication for CNN inference and image filtering─achieving 93.1% CIFAR-10 accuracy under realistic nonidealities─highlight the platform's capability for energy-efficient, beyond-von Neumann computation. The proposed 2D-material-based platform enables heterogeneous functional partitioning while preserving programmability across vertically integrated tiers. The unified 3D architecture supports multimodal neuromorphic operation with layer-level specialization and intertier signal coupling, covalidating learning and system-level functionality within a single integrated stack.