A Preliminary Investigation into Approximating Power Transistor Switching Behavior using a Multilayer Perceptron

Abstract

This paper presents an approach for reproducing key characteristics of non-linear, high frequency switching transients using a multilayer perceptron neural network. Training data is generated using variable time-step transient simulations of a half-bridge switching cell of SPICE transistor models together with constrained yet randomized combinations of DC-link voltage, drain currents and lumped loop inductances. Using the example of peak turn-OFF voltage overshoot for SiC and Si power transistors, the multilayer perceptrons show a mean error of less than (0.9 ± 1.3)%. The predictions of the multilayer perceptron are then compared to preliminary measurements made using a SiC half-bridge test-bench where good agreement is observed especially for higher drain currents. With continued development, such a neural network could be used in coarse, fixed-time-step simulations of any “half-bridge-based” circuit to offer typically unavailable high-fidelity information with negligible computation time. For example, a designer could choose a transistor and quickly see the limits on allowable loop inductance to avoid excessive voltage overshoot for their simulated current waveforms, or see an estimate for voltage overshoot if the loop inductances are known.