Charge-Trap Memory with Engineered Temporal Dynamics for Physically Integrated Reservoir Computing

Area of Science:

• Materials Science
• Artificial Intelligence
• Neuromorphic Computing

Background:

• 2D material (2DM)-based reservoir computing (RC) offers low-power, efficient processing for edge AI.
• Current systems face challenges with volatile memory states and limited retention times.

Purpose of the Study:

• To demonstrate a homogeneous RC system using defect engineering in PdSe2 charge-trap memory (CTM).
• To convert volatile memory states to nonvolatile states for improved performance.
• To advance energy-efficient AI hardware for edge computing and biomedical applications.

Main Methods:

• Utilized ultrafast photoexcitation to induce defect engineering in PdSe2 CTM, creating PdSe2-xOx nanodefects.
• Introduced electron-depleting defects and scattering centers to enhance memory retention.
• Leveraged dual nonlinear/stable operational modes for reservoir computing tasks.

Main Results:

• Achieved conversion from volatile (≈0% retention) to nonvolatile (≈80% retention) states.
• Extended relaxation time constants from 15.6 s to 99.4 s and enabled multilevel memory (>2^6 levels) with prolonged retention (>2000 s).
• Demonstrated high classification accuracy: 91.7% (MNIST) and 93.3% (spoken digits).
• Pioneered electrocardiogram arrhythmia detection with 92.3% accuracy.

Conclusions:

• Established a defect engineering paradigm for material-intrinsic neuromorphic devices.
• The engineered PdSe2 CTM system shows significant potential for energy-efficient AI hardware.
• This approach advances capabilities for biomedical diagnostics and edge computing applications.