This paper develops a novel feedforward neural networks (FFNNs)-based device-level model from physical IGBT model dataset by the proposed artificial neural network (ANN)-aided data-driven IGBT switching transient modeling approach, so that the physics-based IGBT models can be indirectly integrated into FPGA-based real-time simulation of power converters. The main concept is to fit the turn-on / turn-off transient waveforms generated from a physics-based IGBT model by using multiple FFNNs with the same structure but different coefficients. Each FFNN is trained by a dataset covering the transient voltage / current values corresponding to all possible operating conditions at a given discrete time point during transient. All FFNN coefficients are stored on FPGA. By applying the corresponding FFNN coefficients at each simulation time-step, the switching transient waveforms can then be accurately reproduced. The proposed FFNN-based device-level model is designed into two intellectual property (IP) cores at 200 MHz with a fully pipelined structure, which allows the model to authentically reproduce transient waveforms with a 5 ns resolution. A four-phase floating interleaved boost converter (FIBC) is selected as a case study and simulated on a NI-PXIe FlexRIO FPGA real-time platform. The FPGA-based experimental results are compared with that from the LTspice offline simulator, which enables to validate the accuracy and effectiveness of the proposed modeling approach for real-time simulation of power converters.