An FPGA-based Power Converter Simulation Accelerator towards Highly Time-Efficient Machine Learning-Aided Design Methodology

This paper discusses an FPGA-based simulation accelerator for power converters, aiming at achieving a time-efficient machine learning (ML)-aided design methodology. The existing design methodology for power converters is mainly based on human experience and decision, which means that the design results from different designers differ. In addition, this process can be time-consuming due to necessary human-involved design iterations. Recently, ML-aided design methodology has shown its promise to be a great candidate for shortening the circuit design process while guaranteeing high design quality. However, the ML-aided design methodology requires data from a large number of circuit simulations to optimize design parameters. Therefore, simulation time is a crucial bottleneck to further improve the design time-efficiency. This paper proposes a novel FPGA-based circuit simulation tool for fast time-domain simulation. The applied circuit model based on modified nodal analysis (MNA) is briefly described, and a domain-specific architecture (DSA) optimized for FPGA implementation is also explained. The simulation time of conventional and the proposed methods was compared, and a 22.32x speedup was achieved. The proposed design tool based on the proposed time-efficient ML-aided design methodology was applied for designing an LLC resonant converter because its design is challenging and non-unified. The prototype with 330 to 400 V input and 12 V/25 A output was built to verify the proposed design. Importantly, this design methodology can be used for various power electronic circuit topologies, not limited to DC-DC converters. We believe that it holds the potential to drastically shorten the design time through ML-aided design automation and optimization, reduce time-to-market, boost productivity, and accelerate the pace of energy innovation.