How to use AI to achieve inexpensive, fast charging for electric vehicle batteries.
One of the projects we are currently finalizing involves discovering new materials to reduce charging times for EDV batteries. We aim to provide a technological setup for simulating all possible combinations of certain promising materials to develop new synthesized products with the electrical properties necessary for cheap supercharging batteries ( from hours to minutes).
Our setup includes CPUs, GPUs, and FPGAs tailored for running DLA and DFT-based simulations. The programming aspect is quite conventional, using ARM and Risc V assembly for CPUs, VHDL (and some Verilog from the USA) for FPGAs, and a blend of MPI, CUDA, and our own stack with modern C++ (versions 11 to 20) for GPUs.
The topology for our network as is lately the norm for is arranged in a hypercube and enhanced with FPGA SerDes for GPU-GPU and GPU-CPU intercommunications. This setup is key to distributing the DLA particle aggregation process across various GPUs and performing DFT-based electronic structure calculations on materials. Computations are spread across GPUs, necessitating the transfer of wavefunction components or density matrices among GPUs to coherently solve the system’s Hamiltonian.
After the parallel simulations, multiplexing and reduction operations are carried out by the CPUs, which further are overseeing the simulation process and managing I/O operations. Finally, we have employed some AI just to fine-tune simulation parameters and configurations of the setup. The main part was run with ML algorithms to speed up the search by learning from simulation outcomes, minimizing thus the need for extensive manual analysis.
