Machine learning of molecular properties: Locality and active learning

Abstract

In recent years, the machine learning techniques have shown great potent1ial in various problems from a multitude of disciplines, including materials design and drug discovery. The high computational speed on the one hand and the accuracy comparable to that of density functional theory on another hand make machine learning algorithms efficient for high-throughput screening through chemical and configurational space. However, the machine learning algorithms available in the literature require large training datasets to reach the chemical accuracy and also show large errors for the so-called outliers-the out-of-sample molecules, not well-represented in the training set. In the present paper, we propose a new machine learning algorithm for predicting molecular properties that addresses these two issues: it is based on a local model of interatomic interactions providing high accuracy when trained on relatively small training sets and an active learning algorithm of optimally choosing the training set that significantly reduces the errors for the outliers. We compare our model to the other state-of-the-art algorithms from the literature on the widely used benchmark tests.

Improving accuracy of interatomic potentials: more physics or more data? A case study of silica

“…In this work we investigated two strategies for improving the accuracy of a machine-learning interatomic potential, namely adding more fitting parameters to it and adding a charge-equilibration model to it. To that end we tested the MTP potentials [22,23,24] with increasing number of fitting parameters and MTP combined with the charge-equilibration model (MTP+QEq). In order to make a meaningful comparison we assessed the uncertainty of predictions of each potential.…”

“…Another research direction intended to increase the accuracy of the interatomic potentials is the so-called machine-learning interatomic potentials [18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41]. Ideologically, they are different from the empirical interatomic potentials in the way that machine-learning potentials attempt to increase accuracy not by putting more physics into the model, but through a flexible functional form that allows large amounts of DFT data to be used for the fitting.…”

“…In [19] on the basis of the idea of Gaussian process regression, the Gaussian approximation potential (GAP) was proposed and successfully applied for prediction of various properties of carbon, silicon and germanium. Thereafter, many works appeared that propose or validate interatomic potentials based on neural networks [25,26,27,21,28,29,30,31,32,33,34,42], Gaussian processes [20,35,36,37] and other methods [22,23,24,38]. In the above works only short-range interaction was taken into account.…”

Machine-learned multi-system surrogate models for materials prediction

“…11 for single-component system and in Refs. 50,51 was extended to multicomponent systems. MTP partitions the predicted energy into contributions of environments of each atom.…”

Prediction of atomization energy using graph kernel and active learning

“…The chemistry and materials communities have embraced data science and machine learning to bring about revolutionizing solutions to long-standing challenges in molecular modeling, optimal experiment design, and high-throughput structure screening [1][2][3][4][5]. One particularly promising application of machine learning techniques is to train predictive models for molecular properties that are otherwise only available through expensive dynamics simulations and quantum mechanical calculations.…”