Efficient harnessing of polarization dynamics in FeFET-based reservoir computing using split analog method

Reservoir computing (RC) offers an energy-efficient approach to time-series processing by leveraging the inherent dynamics of physical systems. Among various implementations, FeFET-based physical reservoir computing systems are promising owing to the intrinsic nonlinearity and short-term memory of ferroelectric polarization and compatibility with standard CMOS technology. For analog signal applications, the method of converting input signals into gate voltage waveforms is critical to leverage the nonlinearity in ferroelectric polarization dynamics, since ferroelectric polarization dynamics are highly sensitive to applied voltage ranges. In this study, we experimentally examine how different voltage conversion methods influence the performance of FeFET-based RC systems. We first examine a conventional linear analog method, which linearly maps analog inputs to amplitudes of triangular voltage pulses. While simple, this method suffers from significant information loss due to insufficient polarization switching in the low-voltage region, leading to poor computational performance. To overcome this limitation, we propose split analog method, which introduces a voltage gap to exclude non-switching regions, ensuring that input pulses effectively activate ferroelectric dynamics across a wide input range. This method enables access to diverse polarization states by exploiting partial switching behavior. Our experiments demonstrate that split analog method significantly improves performance in input recall, delay, and time-series prediction tasks. Furthermore, combining this method with a gate delay architecture enables high-accuracy prediction even in complex temporal tasks. These results highlight that the careful voltage waveform design transformed to the polarization dynamics is essential for fully leveraging the computational potential of FeFET-based RC systems in neuromorphic applications.