Neuromorphic hardware requires compact, energy-efficient primitives capable of performing neuronal arithmetic and associative computation directly in the analog domain. This thesis investigates Sb–Te–doped GeSe-based Ovonic Threshold Switch (OTS) ...
Neuromorphic hardware requires compact, energy-efficient primitives capable of performing neuronal arithmetic and associative computation directly in the analog domain. This thesis investigates Sb–Te–doped GeSe-based Ovonic Threshold Switch (OTS) devices as the core building blocks for such operations and demonstrates their applicability from material engineering to circuit-level computation. By systematically tuning the Sb–Te content in Ge–Se–Sb–Te (GSST) alloys, a composition-dependent improvement in threshold switching uniformity is identified: the Sb–Te–enriched film Ge31Se31Sb18Te20 exhibits a ~45% reduction in threshold voltage variability compared with a lower-doped counterpart Ge41Se40Sb9Te10. Raman spectroscopy reveals suppression of Se–Se chain modes and consolidation of Ge–Se bonding environments, establishing a direct link between structural homogenization and electrical stability.
Leveraging these stabilized OTS devices, this work implements a set of rate-coded artificial neurons and experimentally demonstrates their arithmetic behavior. The Add neuron, constructed through parallel dendritic branches, exhibits the canonical signatures of additive integration, including preserved I–O slope and full curve collapse when responses are replotted against the summed input rate. A complementary PARALLEL neuron—realized by placing two input branches in series-demonstrates inverse-additive behavior consistent with theoretical expectations, confirming that OTS-based dendritic configuration can reproduce multiple forms of algebraic combination within the unified architectural framework. Extending this principle, three Add neurons with spike-driven lateral inhibition form an Add–Max operator capable of performing the recall phase of a morphological associative memory (MAM). Experiments demonstrate retrieval of all output components and linear mapping between the recalled values and the firing rate of the winning neuron.
A central theme of this thesis is the use of shunt-path modulation as a unifying mechanism across neuronal operators. The same inhibition principle that enables lateral competition in the Add–Max operator is abstracted into a controllable shunt pathway to realize a Division neuron. By introducing a JFET-based, voltage-controlled shunt near the soma, the neuron exhibits divisive gain modulation analogous to biological shunting inhibition. The output firing rate collapses onto a single invariant curve when expressed as the ratio fdr/fmod, and the behavior is accurately modeled by a Hill-type response function. The Division neuron is further applied to pixel-wise image-division, demonstrating compensation of non-uniform illumination.
Overall, this work establishes Sb–Te–doped GeSe OTS devices as a scalable primitive elements capable of supporting biologically inspired arithmetic operations and associative memory computation in compact analog neuromorphic systems.