AP-Net: An Atomic-Pairwise Neural Network for Smooth and Transferable Interaction Potentials

Abstract

<div> <div> <div> <p>Intermolecular interactions are critical to many chemical phenomena, but their accurate computation using <i>ab initio</i> methods is often limited by computational cost. The recent emergence of machine learning (ML) potentials may be a promising alternative. Useful ML models should not only estimate accurate interaction energies, but also predict smooth and asymptotically correct potential energy surfaces. However, existing ML models are not guaranteed to obey these constraints. Indeed, systemic deficiencies are apparent in the predictions of our previous hydrogen-bond model as well as the popular ANI-1X model, which we attribute to the use of an atomic energy partition. As a solution, we propose an alternative atomic-pairwise framework specifically for intermolecular ML potentials, and we introduce AP-Net—a neural network model for interaction energies. The AP-Net model is developed using this physically motivated atomic-pairwise paradigm and also exploits the interpretability of symmetry adapted perturbation theory (SAPT). We show that in contrast to other models, AP-Net produces smooth, physically meaningful intermolecular potentials exhibiting correct asymptotic behavior. Initially trained on only a limited number of mostly hydrogen-bonded dimers, AP-Net makes accurate predictions across the chemically diverse S66x8 dataset, demonstrating significant transferability. On a test set including experimental hydrogen-bonded dimers, AP-Net predicts total interaction energies with a mean absolute error of 0.37 kcal mol−1, reducing errors by a factor of 2-5 across SAPT components from previous neural network potentials. The pairwise interaction energies of the model are physically interpretable, and an investigation of predicted electrostatic energies suggests that the model ‘learns’ the physics of hydrogen-bonded interactions. </p> </div> </div> </div>

Four Generations of High-Dimensional Neural Network Potentials

“…Finally, it should be mentioned that symmetry functions cannot only be constructed to describe atomic energy contributions as a function of the atomic environment according to eq , but also “pair symmetry” functions (PSF) have been suggested, which characterize the geometric environments of pairs of atoms in a similar way , using the energy expression An example for a radial PSF taking into account the environments of both atoms i and j in the pair, which is defined by all atoms k being at least in the cutoff sphere of atom i or j , is the simple function Here, depending on its position, each atom k can thus contribute to one or to both cutoff terms f c ( R ik ) and f c ( R jk ) centered at atoms i and j , respectively. Extensions can be made in analogy to eq by including a Gaussian term yielding Also angular PSFs such as can be constructed.…”

“…Recent comparisons of the orginal ACSFs and the modified forms used in ANI-1 show that both types of functions are equally suited to describe and distinguish atomic environments. 117 Finally, it should be mentioned that symmetry functions cannot only be constructed to describe atomic energy contributions as a function of the atomic environment according to eq 2, but also "pair symmetry" functions (PSF) have been suggested, which characterize the geometric environments of pairs of atoms in a similar way 140,141 using the energy expression…”

Towards chemical accuracy for alchemical free energy calculations with hybrid physics-based machine learning / molecular mechanics potentials

“…Improved interoperability would increase ease of adoption for computational efforts in drug design projects. The definition of the hybrid ML/MM potential could be improved through expanding the terms in the system that are computed with ML by using ML methods such as AIMNet [77], SchNet [78], PhysNet [79], or AP-Net [80] that allow for decomposition of electrostatics and long-range dispersion from short-term valence energies.…”

Development of Range-Corrected Deep Learning Potentials for Fast, Accurate Quantum Mechanical/Molecular Mechanical Simulations of Chemical Reactions in Solution

“…8,16 It is worth mentioning that there are similar procedures being actively pursued in the literature with promising success. [28][29][30][31]33,34,[47][48][49][50][51]55,92 These technologies are still in very early stages, so it is premature to attempt to evaluate their relative merits, but nonetheless should be pointed out and acknowledged. Recent works have used ML training to match semiempirical model forces to ab initio reference data to examine the free-energy surface of glycine condensation.…”