Machine Learning-Based Approximation of MOSFET DC and Small-Signal Characteristics from LTspice Simulation
TL;DR
This study investigates the feasibility of using simple regression-based machine learning models to approximate MOSFET Direct Current and small-signal characteristics obtained from LTspice simulations and shows that polynomial regression achieves very high fitting accuracy.
Abstract
. Accurate modeling of MOSFET characteristics is essential for analog circuit design, where designers commonly rely on SPICE simulations to predict transistor behavior under different bias conditions. However, repeated SPICE simulations can become computationally expensive during design exploration and optimization. This study investigates the feasibility of using simple regression-based machine learning models to approximate MOSFET Direct Current (DC) and small-signal characteristics obtained from LTspice simulations. Linear regression and second-order polynomial regression are used to model the relationships between the gate-to-source voltage (VGS) and two outputs: drain current (Id) and small-signal voltage gain (Av). Results show that polynomial regression achieves very high fitting accuracy. For the Id–VGS relationship, the coefficient of determination is approximately one, indicating that the quadratic model closely matches the simulated data within the studied operating range. For the Av–VGS relationship, the coefficient of determination improves from 0.920 to about 0.990, while the RMSE decreases from 1.53 dB to 0.55 dB.