Spreading huge quantum calculations across many computers

Overview

The problem

Protons are light enough to behave as quantum particles, not just points. Methods that treat protons quantum mechanically alongside the electrons must store an enormous number of particle–particle interactions, and even after compression that data can outgrow a single computer's memory.

The approach

This work spreads the compressed data (Cholesky vectors) across many processors in two stages. Each kind of data is arranged in the layout that lets the key step run without processors constantly exchanging data.

What it found

On water clusters with every proton treated quantum mechanically, the method scales well and keeps the work evenly balanced. One key step for 250 water molecules with 500 quantum protons, the largest calculation of this kind reported, took about 96 seconds on 104 processes. Full calculations on clusters of up to 100 water molecules give a binding energy that extrapolates to a value in excellent agreement with water's measured vaporization enthalpy.

Abstract

Cholesky decomposition (CD) is widely used to reduce the cost of storing two-particle integrals and has been extended to multicomponent nuclear–electronic orbital (NEO) methods. Although NEO-CD lowers memory demands, single-node implementations can still face memory bottlenecks due to the size of the mixed asymmetric nuclear–electronic tensors. In this study, we present a two-stage hybrid-layout distributed NEO-CD framework suitable for large NEO calculations at scale. Cholesky vectors are first generated and orthonormalized using an atomic orbital-index distribution. During Fock matrix construction, mixed nuclear–electronic vectors retain the atomic-orbital-index layout, while pure electronic and pure nuclear vectors are redistributed over the auxiliary index. This design avoids costly inter-rank communication and provides efficient construction of both Coulomb and exchange matrices. Benchmark calculations on water clusters with all protons treated quantum mechanically show excellent scaling and load balance. Using only 104 MPI ranks, a full NEO-Fock build for (H₂O)₂₅₀ with 500 quantum protons, the largest NEO calculation to date, completes in 96.2 s per Fock matrix construction. The distributed NEO-CD framework enables NEO-DFT calculations for liquid-like water clusters up to (H₂O)₁₀₀. Extrapolation of the binding energies yields a bulk-limit value of ΔEₙ→∞ = −10.35 ± 0.19 kcal/mol, in excellent agreement with the experimental vaporization enthalpy of water. These results demonstrate that fully quantum mechanical simulations of large molecular systems are now practical on modern computing architectures.

Nuclear–electronic orbital method at scale: A two-stage hybrid-layout distributed Cholesky decomposition approach

Overview

The problem

The approach

What it found

Abstract