Temporal Adaptivity Enabled High-Efficiency In-Sensor Reservoir Computing Based on MoS₂ Phototransistors.

Area of Science:

• Optoelectronics
• Materials Science
• Computer Science

Background:

• Reservoir computing (RC) is effective for temporal signal processing and efficient hardware development.
• Dynamic target recognition in RC systems is challenged by mismatches in event time scales and optoelectronic properties.

Purpose of the Study:

• To bridge event chronological information with the temporal dynamics of optoelectronic physical nodes in RC.
• To explore the matching of fading memory timescale (τ) with input stimulus time interval (Δt) for improved RC performance.

Main Methods:

• Fabrication of optoelectronic physical nodes using MoS2 phototransistors with varied fading memory timescales (τ).
• Evaluation of linear separability (R²) of reservoir states by varying the ratio of Δt/τ.
• Testing gesture recognition accuracy using a 5-bit optical input.

Main Results:

• Excellent linear separability (R² of 0.988 ± 0.006) achieved when Δt/τ is within 10-20%.
• Gesture recognition accuracy exceeded 85.2% under optimal Δt/τ conditions.
• Lower R² and recognition rates (<77.6%) observed when Δt/τ falls outside the optimal range.

Conclusions:

• Systematic quantification of the relationship between temporal scaling parameters and optical input time intervals is crucial.
• A method for designing temporally adaptive optoelectronic nodes for high-efficiency in-sensor RC systems is provided.