3D-integrated multilayered physical reservoir array for learning and forecasting time-series information.

Area of Science:

• Materials Science
• Computational Neuroscience
• Artificial Intelligence

Background:

• Wide reservoir computing systems offer advanced capabilities for processing multiple time-series data by utilizing parallel reservoir layers.
• Hardware implementation of these systems has been hindered by the lack of high-performance physical reservoirs and complex fabrication processes.

Purpose of the Study:

• To demonstrate a proof-of-principle for a wide physical reservoir computing system.
• To realize efficient learning and forecasting of multiple time-series data using novel hardware.

Main Methods:

• Fabrication of a multilayered, three-dimensional, 3x10x10 tungsten oxide memristive crossbar array.
• Utilizing this 3D stacked array as a physical reservoir for computing tasks.

Main Results:

• Successful proof-of-principle demonstration of a wide physical reservoir computing hardware.
• The three-layer structure effectively extracts intricate 3D local features, outperforming 2D approaches.
• Efficient learning and forecasting of multiple time-series data were achieved.

Conclusions:

• The developed 3D stacked memristive crossbar array enables wide physical reservoir computing.
• This hardware provides a pathway for efficient processing of multiple dynamic time-series information.
• The approach surpasses previous 2D-based methods in extracting complex temporal features.