Fully hardware-oriented physical reservoir computing using 3D vertical resistive switching memory with different...

Area of Science:

• Materials Science
• Computer Science
• Neuromorphic Engineering

Background:

• Reservoir computing (RC) offers a powerful machine learning approach using fixed random networks.
• Current RC implementations often require multiple devices and complex fabrication.
• Integrating reservoir and readout functions is crucial for efficient neuromorphic systems.

Purpose of the Study:

• To develop a fully integrated reservoir computing system.
• To utilize vertical-resistive random-access memory (VRRAM) for both reservoir and readout layers.
• To demonstrate hardware-efficient processing of temporal data and nonlinear system forecasting.

Main Methods:

• A vertically stacked Ta/Ta2O5/HfO2/W and TiN VRRAM structure was designed and fabricated.
• Volatile VRRAM was employed as the physical reservoir, leveraging fading memory and nonlinearity.
• Nonvolatile VRRAM served as the readout network, utilizing multi-level storage and linearity.
• Neuromorphic simulations were conducted to evaluate performance.

Main Results:

• The integrated VRRAM system achieved over 93.14% accuracy in pattern recognition, mimicking biological synapses.
• The Cyclic RC structure demonstrated strong performance in temporal pattern processing.
• Waveform classification achieved an NRMSE of 0.2123, and Hénon map prediction yielded an NRMSE of 0.2377.

Conclusions:

• The proposed VRRAM-based RC system offers a hardware-efficient solution for neuromorphic computing.
• This architecture effectively processes temporal dependencies and forecasts nonlinear dynamical systems.
• The integration of short-term memory functionalities in VRRAM is key for advancing forecasting applications.