Multi-State Conduction in Y-36 Lithium Niobate: Electrical Characterization and NeuroSim Simulation for In-Memory Computing

The computational efficiency of artificial intelligence is becoming more and more constrained by the so called Von Neumann bottleneck causing the transition toward analog Compute-in-Memory (CIM) architectures. This requires synaptic devices capable of combining non-volatile storage, high resolution and as linear as possible conductance modulation. This thesis investigates the potential of ultra-thin (43 nm) Y-36 Lithium Niobate (LiNbO_3) films to address these requirements, through devices characterization and system-level benchmarking. Through electrical characterization of Metal-Ferroelectric-Metal devices an optimized pulse protocol was developed to handle the switching dynamics. This approach was able to set 102 distinct conductance states (more than 6-bit precision) within an analog window operating at currents as low as" " 3" " μ"A" . System-level benchmarking via NeuroSim software on an MNIST set classification task showed that high synaptic resolution and acceptable linearity is necessary for training stability. The optimized configuration achieved an accuracy of 77.61%. Hardware analysis highlighted a performance duality: while the device exhibits remarkable read energy efficiency (~435 µJ), the write energy is currently high due to long pulse duration and high device area. Theoretical projections on future works indicate that overcoming these limits by optimizing switching dynamics to faster regimes can reduce consumption to the femtojoule level, validating Y-36 LiNbO_3 as a scalable platform for next generation neuromorphic computing.