Embedded Files
Machine Learning (ML) is an advanced, modern tool that allows one to build general models of various properties and establish highly non-linear, non-trivial correlations between different characteristics of a system.
An atomistic quantum dynamics simulation requires a classical atomic trajectory, a quantum description of electronic structure, and couplings between different degrees of freedom, such as nonadiabatic electron-vibrational, Coulomb electron-electron, and spin-orbit couplings. These properties can be obtained using supervised ML models, which are based on physically motivated features and are trained to higher level, e.g., ab initio, calculations. Unsupervised ML tools help us to find structure-property correlations that cannot be captured by more traditional regression analyses.
Machine Learning Force Fields enable large-scale and long-time molecular dynamics simulations, many thousands of atoms and nanoseconds, allowing one to sample slow motions and rare events. Rare events are ubiquitous in realistic condensed matter and nanoscale systems, and are associated with various types of defects, amorphous/glassy regions, and multiple components evolving in different timescales.
W. Chu, W. A. Saidi, O. V. Prezhdo, “Long-Lived Hot Electron in a Metallic Particle for Plasmonics and Catalysis: Ab Initio Nonadiabatic Molecular Dynamics with Machine Learning”, ACS Nano, 14, 10608 (2020)
D. Y. Liu, Y. F. Wu, A. S. Vasenko, O. V. Prezhdo, “Grain Boundary Sliding and Distortion on a Nanosecond Timescale Induce Trap States in CsPbBr3: Ab Initio Investigation with Machine Learning Force Field”, Nanoscale, 15, 285-293 (2022)
Machine Learning parameterizations of Density Functional Theory (DFT) enable large-scale electronic structure calculations and provide a general method to obtain DFT parameters for different systems. We developed Nonadiabatic Molecular Dynamics based on ML DFT, making it applicable to thousand atom systems and nanosecond timescales.
D. Liu, B. Wang, Y. Wu, A. S. Vasenko, O. V. Prezhdo, “Breaking the Size Limitation of Non-Adiabatic Molecular Dynamics in Condensed Matter Systems with Local Descriptor Machine Learning”, Proc. Nat. Acad. Sci. USA, 121, e2403497121 (2024)
Nonadiabatic Molecular Dynamics calculations require knowledge of quantum electronic properties along trajectories, including electronic energies, wave functions and nonadiabatic electron-vibrational coupling. Rather than computing these properties quantum mechanically at every timestep, we compute them sparsely along trajectories, e.g., every 64 timesteps, and use ML models to interpolate in-between.
W. B. How, B. P. Wang, W. B. Chu, S. M. Kovalenko, A. Tkatchenko, O. V. Prezhdo, “Dimensionality Reduction in Machine Learning for Nonadiabatic Molecular Dynamics: Effectiveness of Elemental Sublattices in Lead Halide Perovskites”, J. Chem. Phys., 156, 8, 054110 (2022)
B. P. Wang, W. B. Chu, O. V. Prezhdo, “Interpolating Nonadiabatic Molecular Dynamics Hamiltonian with Inverse Fast Fourier Transform”, J. Phys. Chem. Lett., 13, 331-338 (2022)
B. P. Wang, L. Winkler, Y. F. Wu, K. R. Muller, H. E. Sauceda, O. V. Prezhdo, “Interpolating Nonadiabatic Molecular Dynamics Hamiltonian with Bidirectional Long Short-Term Memory Networks”, J. Phys. Chem. Lett., 14 7092-7099 (2023)
Unsupervised learning measures, such as Mutual Information, uncover hidden, non-linear structure-property correlations, helping us to make sense of quantum dynamics data and interpret unusual experimental observations.
G. Q. Zhou, W. Chu, O. V. Prezhdo, “Structural Deformation Controls Charge Losses in MAPbI3: Unsupervised Machine Learning of Nonadiabatic Molecular Dynamics”, ACS Energ. Lett., 5, 1930 (2020)
Page updated
Google Sites
Report abuse