Classical density functional theory in three dimensions with GPU-accelerated automatic differentiation: Computational performance analysis using the example of adsorption in covalent-organic frameworks
Rolf Stierle, Gernot Bauer, Nadine Thiele, Benjamin Bursik, Philipp Rehner, Joachim Groß
Abstract
We show how classical density functional theory can greatly benefit from algorithmic advances in machine learning, especially neural networks. By exploiting GPU-accelerated backward automatic differentiation, we overcome the often cumbersome and error-prone implementation of functional derivatives for classical density functional theory computations. This provides an efficient and straightforward solution for computing functional derivatives, opening up a wide range of applications. We show the gain in computational performance by using backward automatic differentiation to compute the functional derivatives on GPUs, and exemplify the use of this easy-to-implement and highly extensible classical density functional theory framework to predict the adsorption isotherms of a methane/ethane mixture described by a Helmholtz energy functional based on the PC-SAFT equation of state in the covalent-organic framework 2,3-DhaTph. Together with this manuscript, we provide the full classical density functional theory code as supplementary material.