Mixed Precision Fermi-Operator Expansion on Tensor Cores from a Machine Learning Perspective

We present a second-order recursive Fermi-operator expansion scheme using mixed precision floating point operations to perform electronic structure calculations tensor core units. A performance of over 100 teraFLOPs is achieved for half-precision on Nvidia’s A100 The formulated in terms generalized, differentiable deep neural network structure, which solves the quantum mechanical problem. demonstrate how this can be accelerated by optimizing weight and bias values substantially reduce number layers required convergence. also show machine learning approach used optimize coefficients accurately represent fractional occupation numbers states at finite temperatures.