High-Performance Physical Reservoir Computing Based on Phase-Change VO₂ Memristor and Explainable Three-Dimensional...

Area of Science:

• Materials Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Physical reservoir computing (RC) leverages material properties for temporal feature extraction.
• Memristor-based RC offers effective time series processing but faces challenges in explainability and edge compatibility.

Purpose of the Study:

• To develop a novel spiking RC system overcoming limitations of conventional physical RC.
• To utilize phase-change materials for enhanced nonlinear event detection and linear mapping.
• To improve explainability and edge intelligence compatibility in physical RC systems.

Main Methods:

• Constructed a spiking RC system using phase-change materials.
• Implemented a sliding-window spike sampling architecture for device-mapped feature processing.
• Combined simulated annealing with a generative adversarial network for parameter optimization and high-dimensional reservoir creation.
• Introduced a weight-quantification-based explainability analysis for a 3D collaborative mapping mechanism.

Main Results:

• Achieved a 0.075 error rate in Mackey-Glass time series prediction.
• Reached 96.67% accuracy in Iris dataset classification.
• Demonstrated improved explainability through a novel 3D collaborative mapping mechanism.

Conclusions:

• Introduced a novel material system for physical RC, enhancing performance and explainability.
• The developed system shows significant potential for edge intelligence applications.
• The methodology is adaptable to other material platforms, advancing physical RC development.