arXiv:2609.21536
From sparse quantum-computing data to atomistic simulation with universal machine-learning interatomic potentials
Tuan Minh Do, Yuichiro Yoshida, Kenji Ishihara, Wataru Mizukami
Abstract
We propose a framework for incorporating quantum-computing-based electronic-structure calculations into universal machine-learning interatomic potentials (uMLIPs). Rather than constructing an interatomic potential from scratch, we refine a pretrained DFT-based uMLIP using a small set of accurate reference energies obtained from quantum computing. We demonstrate the approach for three chemically distinct applications: the Menshutkin reaction, water adsorption in the metal-organic framework HKUST-1, and CO hopping on a high-entropy-alloy nanoparticle. For the Menshutkin reaction, fine-tuning on gas-phase configurations improves the transition-state energy inside a carbon nanotube but not the product energy. For water adsorption in HKUST-1, fine-tuning with only 14 reference configurations brings adsorption thermodynamics obtained from millions of configurations sampled by Widom insertion into closer agreement with reference values. For CO hopping on an IrPdPtRhRu nanoparticle, the preference for on-top over bridge adsorption is recovered in the finite-temperature free-energy profile obtained from enhanced-sampling molecular dynamics, even though the reference data contain only energies. These results demonstrate that the proposed framework provides a practical route for incorporating quantum-computing calculations into realistic atomistic simulations and that quantum-computing reference data can improve pretrained uMLIPs.
AI Summary
Can sparse quantum-computing electronic energies improve pretrained interatomic potentials enough to support finite-temperature simulations? This paper fine-tunes one MACE universal potential jointly on 26 energy-only configurations: 3 for a Menshutkin reaction, 14 for water adsorption in HKUST-1, and 9 for CO hopping on an alloy nanoparticle. Previously reported reference calculations combine quantum-selected configuration interaction with classical tailored coupled-cluster corrections and, for the adsorption systems, DFT embedding. Quantum active spaces contain only two or four electrons in two or four orbitals. Multihead replay retains energy-and-force training on pretraining data. For the Menshutkin reaction, transferring gas-phase training to carbon-nanotube confinement reduces the barrier error from 15.35 to 4.81 kcal/mol, but worsens the product-energy error. For HKUST-1, Widom insertion over more than two million configurations per run changes the site-I adsorption enthalpy from −28.8 to −47.3 kJ/mol, versus a −48.0 kJ/mol reference. The Henry coefficient moves from $2.0\times10^{-5}$ to $1.1\times10^{-2}$ mol/(kg Pa), versus $1.3\times10^{-2}$, although the reference uncertainty exceeds its central value. For CO hopping, restrained umbrella-sampling dynamics at 300 K reverses the predicted preference from bridge adsorption by 0.19 eV to on-top adsorption by 0.27 eV. Forward/reverse barriers change from 0.34/0.15 to 0.16/0.43 eV; comparison is qualitative against related surface studies rather than a same-system free-energy benchmark. No reference configurations are held out for testing. The paper notes incomplete environmental transfer, possible forgetting during fine-tuning, and sensitivity to training strategy.
Error found: For CO hopping on the alloy nanoparticle, fine-tuning changes the finite-temperature free-energy profile, but its claimed accuracy improvement is not established by a reference calculation or measurement for that system. The HKUST-1 site free-energy expression omits the site-volume factor needed to infer relative populations of the complete site regions, so its interpretation as their relative thermodynamic stability requires an unstated volume-normalized convention.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.