Quantum-centric machine learning for molecular dynamics

TL;DR

The paper tackles the prohibitive cost of ab initio molecular dynamics by eliminating iterative variational optimization in quantum simulations. It introduces quantum-centric machine learning (QCML), a hybrid quantum–classical approach where a Transformer predicts parameterized quantum circuit (PQC) parameters for the variational quantum eigensolver (VQE), enabling direct preparation of molecular wavefunctions $|\Psi(\vec{\theta})\rangle=\hat{U}(\vec{\theta})|\Psi_0\rangle$ and rapid evaluation of energies, forces, and dipole moments. By pretraining on a diverse set of molecules and UCC-based ansatzes and then fine-tuning for new systems, QCML achieves chemical accuracy across potential energy surfaces and spectroscopic observables while delivering orders-of-magnitude speedups over conventional VQE. This framework bridges variational quantum algorithms and modern deep learning, providing a scalable, transferable platform for accurate, efficient quantum chemistry and molecular dynamics simulations, including AIMD and IR spectra predictions. The results demonstrate robust PES fidelity, rapid adaptability to new molecules via fine-tuning, and practical AIMD applications, signaling a new route toward quantum-enabled, data-driven molecular modeling.

Abstract

Accurate and efficient prediction of electronic wavefunctions is central to ab initio molecular dynamics (AIMD) and electronic structure theory. However, conventional ab initio methods require self-consistent optimization of electronic states at every nuclear configuration, leading to prohibitive computational costs, especially for large or strongly correlated systems. Here, we introduce a quantum-centric machine learning (QCML) model-a hybrid quantum-classical framework that integrates parameterized quantum circuits (PQCs) with Transformer-based machine learning to directly predict molecular wavefunctions and quantum observables. By pretraining the Transformer on a diverse dataset of molecules and ansatz types and subsequently fine-tuning it for specific systems, QCML learns transferable mappings between molecular descriptors and PQC parameters, eliminating the need for iterative variational optimization. The pretrained model achieves chemical accuracy in potential energy surfaces, atomic forces, and dipole moments across multiple molecules and ansatzes, and enables efficient AIMD simulations with infrared spectra prediction. This work establishes a scalable and transferable quantum-centric machine learning paradigm, bridging variational quantum algorithms and modern deep learning for next-generation molecular simulation and quantum chemistry applications.

Figures (13)

• Figure 1: Schematic diagram of quantum-centric machine learning (QCML) for molecular dynamics. (a) Pretraining: The variational quantum eigensolver (VQE) algorithm is used to generate the dataset, which is then used to pretrain a Transformer model; (b) Fine-tuning: For new molecules, one can fine-tune the pretrained model using a small dataset, then update the Transformer model that provides an accurate description of ansatz parameters; (c) Property prediction: Given the Transformer model obtained from (b), the electronic wavefunction can be prepared using parameterized quantum circuits and then various molecular properties, such as energies, ionic forces, can be predicted; (d) AIMD: QCML provides highly accurate and efficient evaluations of energies, ionic forces, and dipole moments, enabling long-timescale ab initio molecular dynamics simulations and reliable spectroscopic predictions.
• Figure 1: The non-zero VQE variational parameters of several molecules under different ansatzes.
• Figure 2: Ground-state PESs and their deviation with respect to the FCI results obtained from QCML with different UCC ansatzes for BeH2 (a) and H2O (b). The dotted line indicates chemical accuracy. (c) Comparison between $\Delta \theta$ (the error of the non-zero VQE variational parameters predicted by Transformer model) and $V_{\theta}$ (the sum of the variances of each of the $N_\theta$ non-zero parameters).
• Figure 2: The comparison of loss function with and without gradient clipping.
• Figure 3: Comparison of the average time cost (in second) for computing the single-point energy using VQE and QCML. The black line shows the ratio of the time required for two methods, that is, the speedup of QCML over VQE.