Active Learning of Many-Body Configuration Space: Application to the Cs+–water MB-nrg Potential Energy Function as a Case Study

Abstract

<div> <div> <div> <p>The efficient selection of representative configurations that are used in high-level electronic structure calculations needed for the development of many-body molecular models poses a challenge to current data-driven approaches to molecular simulations. Here, we introduce an active learning (AL) framework for generating training sets corresponding to individual many-body contributions to the energy of a N-body system, which are required for the development of MB-nrg potential energy functions (PEFs). Our AL framework is based on uncertainty and error estimation, and uses Gaussian process regression (GPR) to identify the most relevant configurations that are needed for an accurate representation of the energy landscape of the molecular system under exam. Taking the Cs<sup>+</sup>–water system as a case study, we demonstrate that the application of our AL framework results in significantly smaller training sets than previously used in the development of the original MB-nrg PEF, without loss of accuracy. Considering the computational cost associated with high-level electronic structure calculations for training set configurations, our AL framework is particularly well-suited to the development of many-body PEFs, with chemical and spectroscopic accuracy, for molecular simulations from the gas to condensed phase. </p> </div> </div> </div>

OrbNet: Deep learning for quantum chemistry using symmetry-adapted atomic-orbital features

“…In the context of quantum chemistry, many applications have focused on the use atom-or geometry-specific feature representations and kernel-based [1][2][3][4][5][6][7][8][9] or neural-network (NN) ML architectures. [10][11][12][13][14][15][16][17][18][19][20][21][22][23] Recent studies focus on the featurization of molecules in abstracted representations -such as quantum mechanical properties obtained from low-cost electronic structure calculations [24][25][26][27][28] -and the utilization of novel graphbased neural network [29][30][31][32][33][34][35] techniques to improve transferability and learning efficiency.…”

MB-Fit: Software Infrastructure for Data-Driven Many-Body Potential Energy Functions

“…Future releases of MB-Fit will include an active-learning approach to training set reduction which was shown to be effective in the development of representative training sets for ion-water MB-nrg PEFs. 107,119 C. Quantum mechanical calculations MB-Fit includes an interface that drives QM calculations in order to optimize molecular structures, perform normal-mode analysis, and compute molecular properties (e.g., atomic charges, atomic polarizabilities, and dispersion coefficients). MB-Fit supports running QM calculations locally or, alternatively, provides a job manager that generates short Python scripts for each nB energy calculation which can then be executed on HPC platforms or in a cloud or grid computing environment like Open Science Grid.…”

Geometric Transformer for End-to-End Molecule Properties Prediction

“…Classical molecule property prediction methods combined handcrafted features generally based on force field methods at the atom level (De et al 2016;Faber et al 2017;Christensen et al 2020;Dick and Fernandez-Serra 2020) or molecular level (Rupp et al 2012;Hansen et al 2013;Von Lilienfeld et al 2015;Huang and Von Lilienfeld 2016), integrated into various machine learning models such as Kernel methods (Ramakrishnan et al 2015;Christensen and von Lilienfeld 2019;Zhai et al 2020), Gaussian processes (Bartók et al 2010;Chmiela et al 2017) or Neural networks (Smith, Isayev, and Roitberg 2017;Lubbers, Smith, and Barros 2018;Kearnes et al 2016;Zhang et al 2018;Wu et al 2018;Schütt et al 2017). These methods have recently been superseded by end-toend neural networks, alleviating the need for handcrafted signatures.…”