In the previous paper of this series [JCTC 2020, 16, 3757], we presented a theoretical and algorithmic framework based on a localized representation of the occupied space that exploits the inherent sparsity in the real-space evaluation of the EXX interaction in finite-gap systems. This was accompanied by a detailed description of exx, a massively parallel hybrid MPI/OpenMP implementation of this approach in Quantum ESPRESSO that enables linear-scaling hybrid DFT based AIMD in the NVE/NVT ensembles of condensed-phase systems containing 500--1000 atoms (in fixed orthorhombic cells) with a wall time cost comparable to semi-local DFT. In this work, we extend exx to enable hybrid DFT based AIMD of large-scale condensed-phase systems with general and fluctuating cells in the NpH/NpT ensembles. Our theoretical extension includes an analytical derivation of the EXX contribution to the stress tensor for systems in general cells with a computational complexity that scales linearly with system size. The corresponding algorithmic extensions to exx include optimized routines that: (i) handle static/fluctuating cells with non-orthogonal lattice symmetries, (ii) solve Poissons equation in general cells via an automated selection of the auxiliary grid directions in the Natan-Kronik representation of the discrete Laplacian operator, and (iii) evaluate the EXX contribution to the stress tensor. We also critically assess the computational performance of the extended exx module across several different HPC architectures via case studies on ice Ih, II, and III as well as ambient liquid water. We find that the extended exx can evaluate the EXX contribution to the stress tensor with negligible cost (< 1%) and remains highly scalable, thereby bringing us another step closer to routinely performing hybrid DFT based AIMD for large-scale condensed-phase systems across a wide range of thermodynamic conditions.
Download