A reactive, scalable, and transferable model for molecular energies from a neural network approach based on local information

Abstract

Despite the ever-increasing computer power, accurate ab initio calculations for large systems (thousands to millions of atoms) remain infeasible. Instead, approximate empirical energy functions are used. Most current approaches are either transferable between different chemical systems, but not particularly accurate, or they are fine-tuned to a specific application. In this work, a data-driven method to construct a potential energy surface based on neural networks is presented. Since the total energy is decomposed into local atomic contributions, the evaluation is easily parallelizable and scales linearly with system size. With prediction errors below 0.5 kcal mol for both unknown molecules and configurations, the method is accurate across chemical and configurational space, which is demonstrated by applying it to datasets from nonreactive and reactive molecular dynamics simulations and a diverse database of equilibrium structures. The possibility to use small molecules as reference data to predict larger structures is also explored. Since the descriptor only uses local information, high-level ab initio methods, which are computationally too expensive for large molecules, become feasible for generating the necessary reference data used to train the neural network.

Thermal activation of methane by MgO+: temperature dependent kinetics, reactive molecular dynamics simulations and statistical modeling

“…The NN constructs a descriptor vector for each atom which encodes information about the local chemical environment of the atoms. 33 The total energy of the system is obtained by combining 'atomic energy contributions' predicted from these descriptors. The NN is 10 layers deep with 64 neurons per layer.…”

Reactive molecular dynamics: From small molecules to proteins

“…Using a strictly local chemical descriptor, a NN-based method tailored for accurate energy evaluations, which can be applied to construct PESs for nonreactive and reactive dynamics of chemically heterogeneous systems in the condensed phase, has been introduced . 233 Such a NN trained on 100 k reference structures can learn to accurately predict energies of structures in the QM9 data set 232 across chemical space with a MAE of 0.41 kcal/mol which is only slightly worse than that of the SchNet architecture 231 (MAE of 0.34 kcal/mol). Contrary to SchNet, this NN is considerably more efficient because a local descriptor is used and the network architecture is much simpler.…”

High-dimensional potential energy surfaces for molecular simulations: from empiricism to machine learning

“…Two variants of HDNNs can be distinguished: the 'descriptor-based' variant uses a hand-crafted descriptor [59,[94][95][96], to encode the environment of an atom, which is then used as input of a multi-layer feed-forward NN. Examples for this kind of approach are the 'Accurate NeurAl networK engINe for Molecular Energies' (ANAKIN-ME or ANI) [97] or TensorMol [98].…”