物理的に統合されたリザーバーコンピューティングのための時間ダイナミクスを設計したチャージトラップメモリ

2D material (2DM)-based reservoir computing (RC) systems combine the advantages of low-power hardware implementation with lightweight neural network architectures capable of processing complex temporal patterns through minimal training overhead, positioning them as ideal platforms for edge artificial intelligence (AI) applications. Here, a homogeneous RC system via defect engineering in PdSe₂ charge-trap memory (CTM) by ultrafast photoexcitation is demonstrated, which directly generates PdSe_2-xO_x nanodefects, converting volatile states (≈0% retention) into nonvolatile states (≈80% retention) by introducing electron-depleting defects and scattering centers in PdSe₂ channel. This engineering extends relaxation time constants from 15.6 s to 99.4 s and enables multilevel memory (>2⁶ levels) with prolonged retention (>2000 s). Leveraging dual nonlinear/stable operational modes, the physically integrated RC system achieves 91.7% (MNIST) and 93.3% (spoken digits) classification accuracy. Notably, it pioneers electrocardiogram arrhythmia detection (N, L, R, A, and V classes) with 92.3% accuracy, surpassing existing in-memory computing approaches. By establishing a defect engineering paradigm for material-intrinsic neuromorphic devices, this work advances energy-efficient AI hardware for biomedical diagnostics and edge computing applications.