Abstract
In‐materia reservoir computing is typically benchmarked in segmented regimes where implicit re‐initialization masks long‐term state accumulation. Under continuous, nonstationary drive, however, multi‐timescale device dynamics can retain activity‐dependent residuals that fail to relax to a reproducible baseline, leading to drift and accuracy loss. We introduce reset‐aware physical reservoir computing, where electrical reset is promoted to an addressable, input‐decoupled system operation that can be sparsely triggered by online drift statistics. Trap‐engineered Schottky silicon nanowire FETs (SiNW‐FETs) realize this concept: interfacial trap kinetics and Schottky‐barrier modulation yield multi‐timescale fading memory, while source–drain polarity reversal provides a device‐native erase path independent of gate updates. Using the in‐plane solid–liquid–solid (IPSLS) growth mechanism, aligned SiNW arrays with diameters of ∼20 nm are obtained and integrated into a complementary metal‐oxide semiconductor (CMOS) compatible, addressable 4×4 top‐gated SiNW‐FET array. On a standard event‐based gesture‐recognition benchmark under continuous streaming inference, sparse reverse‐bias erase suppresses drift and sustains >90% accuracy, whereas erase‐free streaming drops abruptly by ∼20% during prolonged operation. These results establish a CMOS‐compatible in‐materia reservoir with explicit, controllable state management for robust continuous event‐driven inference.
Funding Information
  • National Key Research and Development Program of China (2024YFA1208900)
  • Natural Science Foundation of Jiangsu Province (BK20251214)
  • National Natural Science Foundation of China (62325403)
  • National Natural Science Foundation of China (62504103)

This publication has 42 references indexed in Scilit: