Quantum Monte Carlo generation of machine learning potentials for hydrogen-basedhigh-temperature superconductors

• Name: Quantum Monte Carlo generation of machine learning potentials for hydrogen-basedhigh-temperature superconductors
• EuroHPC machine used: Leonardo
• Topic: Natural sciences; physical sciences.

Overview of the project

Hydrogen-based compounds have been in the spotlight of the condensed matter community since the discovery of high-temperature superconductivity in sulfur hydride under high pressure. This has been followed by other findings, revealing an entire family of hydrogen-based compounds that share the desirable property of superconducting at very high temperature, breaking all previous records. The possibility of increasing the critical temperature by playing with hydrogen chemistry opened the door to a new fertile research ground, in the quest for higher superconducting temperatures and lower (and so more technologically exploitable) pressures. From the theory viewpoint, this quest is challenged by the peculiar properties of hydrogen-based superconductors: strong nuclear quantum effects, large phonon anharmonicity, structural near degeneracy, competing energy scales between electrons and nuclei. In this project, we will tackle this problem by combining accurate quantum Monte Carlo (QMC) electronic energies and nuclear forces with an advanced treatment of lattice vibrations, with full inclusion of quantum anharmonicity. This will be achieved by the generation of machine learning potentials, based on QMC-quality training sets, which will then be used in path integral molecular dynamics calculations to simulate lattice dynamics, in targeted compounds, such as pristine hydrogen and the most promising hydrides.

How did EPICURE support the project and what were the benefits of the support?

“The project aimed to exploit the quantum Monte Carlo accuracy to produce reliable training sets for machine learning interatomic potentials (MLIP) generation for hydrogen-rich compounds. This family of compounds is critical in that both accurate MLIP and sophisticated molecular dynamics (MD) simulations, such as path integral MD, are needed to reach a predictive quantum simulation. The EPICURE support has been useful to reduce the cost of single point calculations, allowing the generation of more extensive training sets.

We received assistance to improve the performances of the TurboRVB code run on the Leonardo Booster GPU-accelerated partition. The support activity had two main goals: the first one has been meant to make the GPU usage more efficient, the second one was focused on some software development in the GPU-accelerated part of the code such as the code could be compiled with the latest NVidia Fortran compiler, hopefully delivering better performances by the compiled executable.

The outcome of the collaboration has been twofold. On one side, the usage of the GPU acceleration has largely been improved thanks to the NVIDIA multi-process service (MPS). This GPU sharing solution allows multiple MPI processes (each one accelerated by a single GPU) to share a single physical NVIDIA GPU hardware attached to a node. Without MPS, we were bound to run only 4 MPI processes per node (equipped with 4 GPU). With MPS, we have been able to run up to 32 MPI processes per node (this optimal number depends on the GPU load). With MPS, the measured speed-up has been in the range of 4-6 times faster than without MPS. On the other side, EPICURE support tried to improve the CUDA interface coded in TurboRVB, to make it compatible with the latest Fortran NVIDIA compiler. While able to compile with the latest version, the code performance increase has not been significant. However, the code development brought about the nice feature of being compliable with the most recent NVIDIA compiler, improving its portability.

The results on the performance gain through NVIDIA MPS have been published in the following white paper: Maximizing GPU utilization with NVIDIA Multi-Process-Service for Quantum Monte Carlo simulations with the TurboRVB code, L Bellentani, M Casula, T Gorni. ” –  Michele Casula

Additional references

Load-balanced diffusion Monte Carlo method with lattice regularization, K Nakano, SSorella, M Casula, The Journal of Chemical Physics 163, 194117 (2025);

Self-consistency error correction for accurate machine learning potentials from variational Monte Carlo, G Tenti, K Nakano, M Casula, Journal of Chemical Theory and Computation 21, 9335-9346 (2025);

Displacive Quantum Critical Point in Superconducting Hydrides: The Case of H3S, M Cherubini, A Raghav, M Casula, Physical Review Letters 137, 046102 (2026); Quadrupolar phase transition in superconducting lanthanum hydride, A Raghav, K Nakano, M Cherubini, R Arita, M Casula, arXiv preprint arXiv:2608.10428