The authors introduce an end-to-end active learning (AL) protocol for constructing machine learning potentials (MLPs) that can greatly accelerate quantum chemical simulations. The key aspects of the protocol are:
Physics-informed sampling: The protocol uses different amounts of physical information (energies and gradients) about the potential energy surface (PES) to guide the sampling of training points. This ensures the sampled points capture the important features of the PES.
Automatic initial data selection: The authors propose a method to automatically determine the size of the initial data set based on the expected performance of the MLP. This avoids the need for manual experimentation.
Uncertainty quantification (UQ): The protocol uses the deviation between two MLP models (one trained on energies and gradients, the other on energies only) as the UQ criterion to identify regions of the PES that require further sampling.
The authors demonstrate the effectiveness of this protocol in three applications: vibrational spectra simulations, conformer search, and time-resolved mechanism investigation of the Diels-Alder reaction. In all cases, the protocol was able to construct accurate MLPs with a significantly reduced computational cost compared to direct quantum chemical calculations.
The key advantages are the data-efficiency, robustness, and seamless integration of all the required steps (sampling, labeling, and machine learning) in a single workflow. This enables the authors to break through the bottleneck of expensive molecular dynamics simulations and make them feasible on commodity hardware.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Yi-Fan Hou,L... at arxiv.org 04-19-2024
https://arxiv.org/pdf/2404.11811.pdfDeeper Inquiries