2024 AIChE Annual Meeting

(11k) Ab Initio Molecular Dynamics with Quantum Hardware and Transfer Learning

Authors

Batton, C. - Presenter, University of California, Berkeley
Khan, A., University of Illinois Urbana-Champaign
Vaish, P., Brown University
Pang, Y., Brown University
Chen, M. S., New York University
Rotskoff, G. M., Stanford University
Mullinax, J. W., NASA Ames Research Center
Rubenstein, B., Brown University
Tubman, N., ASA Ames Research Center
The ability to perform ab initio molecular dynamics with both high accuracy and long timescales would allow for the precise determination of thermodynamic and kinetic properties in such processes as catalysis and biophysics. Here, we present a general framework to do so utilizing neural network potentials and quantum hardware. Fitting neural network potentials to energetics and forces obtained through quantum hardware would allow for such highly accurate and fast simulations. However, this process is limited by the current noisy energies computed on quantum hardware, along with the difficulty in scaling to large systems.

To alleviate the current difficulties in quantum hardware we utilize transfer learning, in which a neural network potential is trained initially on lower-quality data produced using classical computers and then fine-tuned on high-quality data generated on quantum hardware. Here, an initial neural network potential is fit to data from density functional theory, which is used to probe the behavior in the system of interest. To generate the configurations in the quantum hardware dataset, we use an active learning procedure to identify informative configurations for the subsequent refining of the neural network potentials. High accuracy results for this dataset are then obtained using the variational quantum eigensolver. This approach significantly reduces the quantum training dataset size while maintaining high accuracy of the potentials, allowing us to obtain accurate thermodynamic and kinetic information about model reactive systems.