NNAIMQ: A neural network model for predicting QTAIM charges

Abstract

Atomic charges provide crucial information about the electronic structure of a molecular system. Among the different definitions of these descriptors, the one proposed by the Quantum Theory of Atoms in Molecules (QTAIM) is particularly attractive given its invariance against orbital transformations although the computational cost associated with their calculation limits its applicability. Given that Machine Learning (ML) techniques have been shown to accelerate orders of magnitude the computation of a number of quantum mechanical observables, in this work, we take advantage of ML knowledge to develop an intuitive and fast neural network model (NNAIMQ) for the computation of QTAIM charges for C, H, O, and N atoms with high accuracy. Our model has been trained and tested using data from quantum chemical calculations in more than 45 000 molecular environments of the near-equilibrium CHON chemical space. The reliability and performance of NNAIMQ have been analyzed in a variety of scenarios, from equilibrium geometries to molecular dynamics simulations. Altogether, NNAIMQ yields remarkably small prediction errors, well below the 0.03 electron limit in the general case, while accelerating the calculation of QTAIM charges by several orders of magnitude.

Open-Source Machine Learning in Computational Chemistry

“…150 To predict Quantum Theory of Atoms in Molecules' PACs, the NNAIMQ (an NN model) was created. 151 The SuperAtomicCharge model, a feed-forward NN, was written to predict QM-derived RESP, DDEC4, and DDEC78 PACs. 152 For metal-containing systems, mpn_charges 153 and PAC in Metal−Organic Frameworks (PACMOF) 154 were developed using message-passing NN and a random-forest approach, respectively.…”

“…Focusing on small molecules, the Atom-Path-Descriptor (APD) uses a new type of atomic descriptor for training random forest and extreme gradient boosting models for predicting PAC . To predict Quantum Theory of Atoms in Molecules’ PACs, the NNAIMQ (an NN model) was created . The SuperAtomicCharge model, a feed-forward NN, was written to predict QM-derived RESP, DDEC4, and DDEC78 PACs .…”

Developing a User-Friendly Code for the Fast Estimation of Well-Behaved Real-Space Partial Charges

“…• Model parameters: Specify the main ML kernel used to obtain the raw atomic properties in the first place. Both builtin (NNAIMQ 13 ) and tailor-made models can be used. If custom kernels are employed, which can be straightforwardly trained with NNAIMGUI, the path to the model folder (Model) must be given along with the name of the target property (Target prop) and property units (Units).…”

New venues in electron density analysis

“…Within this context, the development of Neural Networks (NNs), able to approximate any function with arbitrary accuracy, 43,44 is reshaping the course of quantum and computational chemistry. Indeed, the blatant success of Deep Learning (DL) and related strategies in the chemistry realm has led to faster, yet reliable, tools for many different purposes such as property prediction, 36,45–49 atomistic simulations, 50,51 quantum-chemically accurate force fields 52–54 and potentials 55,56 or novel techniques for chemical discovery and sampling, 57–59 to name just a few.…”

“…Finally, in passing it may also be worth mentioning that beyond energetic partitioning schemes, QTAIM in general can also benefit from the computational boost offered by ML techniques. Indeed, some of us have worked 45 in a NN model capable of predicting the local value of the electron density within a topological atom, in the form of atomic charges. The success of these and similar NN models suggests that the boundaries of QCT in chemistry are likely to vanish in the very near future owing to the remarkable performance of ML-driven quantum chemical tools.…”