Scott Tan1 Shinhyun Choi1 Zefan Li1 Yunjo Kim1 Chanyeol Choi1 Pai-Yu Chen2 Hanwool Yeon1 Shimeng Yu2 Jeehwan Kim1

1, Massachusetts Institute of Technology, Cambridge, Massachusetts, United States
2, Arizona State University, Tempe, Arizona, United States

New platforms specialized for artificial intelligence (AI) are necessary to reduce power consumption of deep neural network operations. One potential hardware alternative is a neuromorphic computing system, which utilizes a crossbar structure comprised of artificial synapses at crosspoints for fast synaptic weight update and low-power vector-matrix multiplications. Each array crosspoint must be capable of accessing many analog conductance states by symmetric potentiation and depression in response to uniform voltage pulses.

We have developed epitaxial random-access memory (epiRAM) as a hardware platform for neuromorphic computing. As a switching medium, single-crystalline Silicon (Si) prevents reduction of Silver ions (Ag+) within the lattice, which prohibits conductive bridging at typical operating temperatures. However, we have found that epitaxial Silicon-Germanium (Si-Ge) is a suitable solid electrolyte for Ag+ migration and Ag conduction channel evolution through threading dislocation pathways. Heteroepitaxial Si-Ge films are grown on Si in the kinetically-limited metastable regime to suppress dislocation glide. Threading pathways are widened using defect-selective etching. Within artificial synapses, metallic conduction pathways form through widened threading dislocations in the Si-Ge layer. This channel evolution allows artificial synapses to have tunable conductance states. Using metastable Si-Ge, we demonstrate that analog resistive switching can occur through devices with electrode area smaller than 25 x 25 nm with similar Current-Voltage (I-V) characteristics to bigger (5 µm x 5 µm) devices. This result suggests channel evolution is dominated by localized conductive regions comprised of dense threading dislocation arrays.

Synaptic weights in neural networks can be updated and stored as conductance states of epiRAM devices at array crosspoints. Conduction channel evolution is confined for linear and gradual conductance change during neural network training. Switching threshold voltage and read current tuning is accomplished by adjusting the Schottky barrier at the p+ Si/ Ag conduction channel interface. Semiconductor doping can rectify sneak currents to enable large-scale arrays. EpiRAM in passive crossbars could be trained for MNIST handwriting digit recognition with up to 95.1% accuracy, only 2% lower than the binary Multilayer Perceptron (MLP) software baseline. Thus, epitaxial Si-Ge artificial synapses are a promising new hardware platform for AI.