In this work, we aim to unveil the general correlations between the performance of a physical reservoir computing (RC) system and the inherent nonlinear dynamics of the adopted device. Taking the metal-ferroelectric-metal (MFM) capacitor, one of the most popular candidate devices for compute-in-memory (CIM) technology, as the computational platform, we construct a nonlinear dynamical model of polarization in the ferroelectric layer. We then design the physical RC utilizing a single and/or an array of MFM capacitors by analyzing the model's stability and feasible dynamical cases. Subsequently, both the initial task and benchmark are numerically conducted to verify the designed RC's superiority. It is proven that by selecting an appropriate dynamical case, the RC can achieve a recognition rate as high as 96.13%, surpassing the results reported in previous work. Finally, we discuss how these key parameters play their role in the RC's performance from the perspective of affecting the system's transient responses, nonlinearity, and short-fading memory. This work paves the foundation for designing highly efficient reservoir computing based on MFM capacitors as well as other memristive devices such as memristors, tunneling diodes, etc.
