This is project aims at accelerate the circuit simulation process when using machine learining to train the power converter designer. The algorithm is based on AutoCkt ¹². An FPGA accelerator is implemneted via xilinx vitis to accelerate the circuit simulation. Half-bridge LLC converter is used as the example. Compare with using PowerSim as the simulator, it is 60x fatser. This project contains:

• A python script that transfer spice like netlist into accelerator required matrixes and vectors.
• A Xilinx Vitis project to create the hardware accelerator.
• The training and forwarding script that used to train the circuit designer. A customed GYM environment that use FPGA accelerator to do the simulation.

To run the project, an AWS F1 instance is required. The following python packages (and there dependencies) are required:

• ray
• ray[rllib]
• pynq

These three packages may not be installed correctly with conda. If so, please use python3 -m pip install <package name>

To run the training, run following teriminal commands.

git clone https://github.com/aws/aws-fpga.git
cd aws-fpga
source vitis_setup.sh
cd ..
git clone https://github.com/URI-nextlab/ML-Powelec.git
cd ML-Powelec/fpga_train
ipython
run train.py