Abstract
Reservoir computing (RC) offers advantages in processing time-series data with reduced training costs and simpler architectures. This study presents a hardware-implemented RC system utilizing multifunctional memristors fabricated using a single process. By leveraging a ferroelectric-based memristor (FM) as a volatile reservoir layer and a redox-based memristor (RM) as a non-volatile readout layer, seamless integration without additional fabrication steps is achieved. The dual-functional memristor structure enables electrical conversion from FM to RM, enhancing system scalability and versatility. Comprehensive electrical measurements, including low-frequency noise analysis and weight update linearity evaluation, validate the memristors' performance. Potentiation and depression processes achieve a linearity factor improvement to ensure precise synaptic weight tuning, with cycle-to-cycle variation <2.3%. Additionally, the ferroelectric-based memristor exhibits a cycle-to-cycle variation of 3.52%, maintaining distinct reservoir states with minimal overlap. Offline training demonstrates a high classification accuracy of 93.3% on the Modified National Institute of Standards and Technology dataset, while online training achieves an accuracy of 88.2% with incremental pulse schemes, surpassing the accuracy of identical pulse schemes (65.1%). These results establish the practical viability of multifunctional memristors for neuromorphic systems, establishing a robust foundation for next-generation computing technologies.