Reducing the Cost of Neural Network Potential Generation for Reactive Molecular Systems

Efficient exploration of reaction pathways using reaction databases and active learning

The fast and accurate simulation of chemical reactions is a major goal of computational chemistry. Recently, the pursuit of this goal has been aided by machine learning interatomic potentials (MLIPs), which provide energies and forces at quantum mechanical accuracy but at a fraction of the cost of the reference quantum mechanical calculations. Assembling the training set of relevant configurations is key to building the MLIP. Here, we demonstrate two approaches to training reactive MLIPs based on reaction pathway information. One approach exploits reaction datasets containing reactant, product, and transition state structures. Using an SN2 reaction dataset, we accurately locate reaction pathways and transition state geometries of up to 170 unseen reactions. In another approach, which does not depend on data availability, we present an efficient active learning procedure that yields an accurate MLIP and converged minimum energy path given only the reaction end point structures, avoiding quantum mechanics driven reaction pathway search at any stage of training set construction. We demonstrate this procedure on an SN2 reaction in the gas phase and with a small number of solvating water molecules, predicting reaction barriers within 20 meV of the reference quantum chemistry method. We then apply the active learning procedure on a more complex reaction involving a nucleophilic aromatic substitution and proton transfer, comparing the results against the reactive ReaxFF force field. Our active learning procedure, in addition to rapidly finding reaction paths for individual reactions, provides an approach to building large reaction path databases for training transferable reactive machine learning potentials.

Random Sampling Versus Active Learning Algorithms for Machine Learning Potentials of Quantum Liquid Water

Training accurate machine learning potentials requires electronic structure data comprehensively covering the configurational space of the system of interest. As the construction of this data is computationally demanding, many schemes for identifying the most important structures have been proposed. Here, we compare the performance of high-dimensional neural network potentials (HDNNPs) for quantum liquid water at ambient conditions trained to data sets constructed using random sampling as well as various flavors of active learning based on query by committee. Contrary to the common understanding of active learning, we find that for a given data set size, random sampling leads to smaller test errors for structures not included in the training process. In our analysis, we show that this can be related to small energy offsets caused by a bias in structures added in active learning, which can be overcome by using instead energy correlations as an error measure that is invariant to such shifts. Still, all HDNNPs yield very similar and accurate structural properties of quantum liquid water, which demonstrates the robustness of the training procedure with respect to the training set construction algorithm even when trained to as few as 200 structures. However, we find that for active learning based on preliminary potentials, a reasonable initial data set is important to avoid an unnecessary extension of the covered configuration space to less relevant regions.

ArcaNN: automated enhanced sampling generation of training sets for chemically reactive machine learning interatomic potentials

The emergence of artificial intelligence is profoundly impacting computational chemistry, particularly through machine-learning interatomic potentials (MLIPs). Unlike traditional potential energy surface representations, MLIPs overcome the conventional computational scaling limitations by offering an effective combination of accuracy and efficiency for calculating atomic energies and forces to be used in molecular simulations. These MLIPs have significantly enhanced molecular simulations across various applications, including large-scale simulations of materials, interfaces, chemical reactions, and beyond. Despite these advances, the construction of training datasets-a critical component for the accuracy of MLIPs-has not received proportional attention, especially in the context of chemical reactivity, which depends on rare barrier-crossing events that are not easily included in the datasets. Here we address this gap by introducing ArcaNN, a comprehensive framework designed for generating training datasets for reactive MLIPs. ArcaNN employs a concurrent learning approach combined with advanced sampling techniques to ensure an accurate representation of high-energy geometries. The framework integrates automated processes for iterative training, exploration, new configuration selection, and energy and force labeling, all while ensuring reproducibility and documentation. We demonstrate ArcaNN's capabilities through two paradigm reactions: a nucleophilic substitution and a Diels-Alder reaction. These examples showcase its effectiveness, the uniformly low error of the resulting MLIP everywhere along the chemical reaction coordinate, and its potential for broad applications in reactive molecular dynamics. Finally, we provide guidelines for assessing the quality of MLIPs in reactive systems.

Introduction to machine learning potentials for atomistic simulations

Machine learning potentials have revolutionised the field of atomistic simulations in recent years and are becoming a mainstay in the toolbox of computational scientists. This paper aims to provide an overview and introduction into machine learning potentials and their practical application to scientific problems. We provide a systematic guide for developing machine learning potentials, reviewing chemical descriptors, regression models, data generation and validation approaches. We begin with an emphasis on the earlier generation of models, such as high-dimensional neural network potentials and Gaussian approximation potentials, to provide historical perspective and guide the reader towards the understanding of recent developments, which are discussed in detail thereafter. Furthermore, we refer to relevant expert reviews, open-source software, and practical examples-further lowering the barrier to exploring these methods. The paper ends with selected showcase examples, highlighting the capabilities of machine learning potentials and how they can be applied to push the boundaries in atomistic simulations.

Surprising Dynamics Phenomena in the Diels-Alder Reaction of C60 Uncovered with AI

We performed an extensive artificial intelligence-accelerated quasi-classical molecular dynamics investigation of the time-resolved mechanism of the Diels-Alder reaction of fullerene C60 with 2,3-dimethyl-1,3-butadiene. In a substantial fraction (10%) of reactive trajectories, the larger C60 noncovalently attracts the 2,3-dimethyl-1,3-butadiene long before the barrier so that the diene undergoes the series of complex motions including roaming, somersaults, twisting, and twisting somersaults around the fullerene until it aligns itself to pass over the barrier. These complicated processes could be easily missed in typically performed quantum chemical simulations with shorter and fewer trajectories. After the barrier is passed, the bonds take longer to form compared to the simplest prototypical Diels-Alder reaction of ethene with 1,3-butadiene despite high similarities in transition states and barrier widths evaluated with intrinsic reaction coordinate (IRC) calculations. C60 is mainly responsible for these differences as its reaction with 1,3-butadiene is similar to the reaction with 2,3-dimethyl-1,3-butadiene: the only substantial difference being that the extra methyl groups double the probability of the prolonged alignment phase in dynamics. These additional calculations of C60 with 1,3-butadiene could be performed via active learning more easily by reusing the data generated for the other two reactions, showing the potential for larger-scale exploration of the effects of different substrates in the same types of reactions.

Physics-Informed Active Learning for Accelerating Quantum Chemical Simulations

Quantum chemical simulations can be greatly accelerated by constructing machine learning potentials, which is often done using active learning (AL). The usefulness of the constructed potentials is often limited by the high effort required and their insufficient robustness in the simulations. Here, we introduce the end-to-end AL for constructing robust data-efficient potentials with affordable investment of time and resources and minimum human interference. Our AL protocol is based on the physics-informed sampling of training points, automatic selection of initial data, uncertainty quantification, and convergence monitoring. The versatility of this protocol is shown in our implementation of quasi-classical molecular dynamics for simulating vibrational spectra, conformer search of a key biochemical molecule, and time-resolved mechanism of the Diels-Alder reaction. These investigations took us days instead of weeks of pure quantum chemical calculations on a high-performance computing cluster.

Prebiotic chemical reactivity in solution with quantum accuracy and microsecond sampling using neural network potentials

While RNA appears as a good candidate for the first autocatalytic systems preceding the emergence of modern life, the synthesis of RNA oligonucleotides without enzymes remains challenging. Because the uncatalyzed reaction is extremely slow, experimental studies bring limited and indirect information on the reaction mechanism, the nature of which remains debated. Here, we develop neural network potentials (NNPs) to study the phosphoester bond formation in water. While NNPs are becoming routinely applied to nonreactive systems or simple reactions, we demonstrate how they can systematically be trained to explore the reaction phase space for complex reactions involving several proton transfers and exchanges of heavy atoms. We then propagate at moderate computational cost hundreds of nanoseconds of a variety of enhanced sampling simulations with quantum accuracy in explicit solvent conditions. The thermodynamically preferred reaction pathway is a concerted, dissociative mechanism, with the transient formation of a metaphosphate transition state and direct participation of water solvent molecules that facilitate the exchange of protons through the nonbridging phosphate oxygens. Associative-dissociative pathways, characterized by a much tighter pentacoordinated phosphate, are higher in free energy. Our simulations also suggest that diprotonated phosphate, whose reactivity is never directly assessed in the experiments, is significantly less reactive than the monoprotonated species, suggesting that it is probably never the reactive species in normal pH conditions. These observations rationalize unexplained experimental results and the temperature dependence of the reaction rate, and they pave the way for the design of more efficient abiotic catalysts and activating groups.

A personal perspective of the present status and future challenges facing thermal reaction rate theory

Reaction rate theory has been at the center of physical chemistry for well over one hundred years. The evolution of the theory is not only of historical interest. Reliable and accurate computation of reaction rates remains a challenge to this very day, especially in view of the development of quantum chemistry methods, which predict the relevant force fields. It is still not possible to compute the numerically exact rate on the fly when the system has more than at most a few dozen anharmonic degrees of freedom, so one must consider various approximate methods, not only from the practical point of view of constructing numerical algorithms but also on conceptual and formal levels. In this Perspective, I present some of the recent analytical results concerning leading order terms in an ℏ2m series expansion of the exact rate and their implications on various approximate theories. A second aspect has to do with the crossover temperature between tunneling and thermal activation. Using a uniform semiclassical transmission probability rather than the "primitive" semiclassical theory leads to the conclusion that there is no divergence problem associated with a "crossover temperature." If one defines a semiclassical crossover temperature as the point at which the tunneling energy of the instanton equals the barrier height, then it is a factor of two higher than its previous estimate based on the "primitive" semiclassical approximation. In the low temperature tunneling regime, the uniform semiclassical theory as well as the "primitive" semiclassical theory were based on the classical Euclidean action of a periodic orbit on the inverted potential. The uniform semiclassical theory wrongly predicts that the "half-point," which is the energy at which the transmission probability equals 1/2, for any barrier potential, is always the barrier energy. We describe here how augmenting the Euclidean action with constant terms of order ℏ2 can significantly improve the accuracy of the semiclassical theory and correct this deficiency. This also leads to a deep connection with and improvement of vibrational perturbation theory. The uniform semiclassical theory also enables an extension of the quantum version of Kramers' turnover theory to temperatures below the "crossover temperature." The implications of these recent advances on various approximate methods used to date are discussed at length, leading to the conclusion that reaction rate theory will continue to challenge us both on conceptual and practical levels for years to come.

Screening Cu-Zeolites for Methane Activation Using Curriculum-Based Training

Machine learning (ML), when used synergistically with atomistic simulations, has recently emerged as a powerful tool for accelerated catalyst discovery. However, the application of these techniques has been limited by the lack of interpretable and transferable ML models. In this work, we propose a curriculum-based training (CBT) philosophy to systematically develop reactive machine learning potentials (rMLPs) for high-throughput screening of zeolite catalysts. Our CBT approach combines several different types of calculations to gradually teach the ML model about the relevant regions of the reactive potential energy surface. The resulting rMLPs are accurate, transferable, and interpretable. We further demonstrate the effectiveness of this approach by exhaustively screening thousands of [CuOCu]2+ sites across hundreds of Cu-zeolites for the industrially relevant methane activation reaction. Specifically, this large-scale analysis of the entire International Zeolite Association (IZA) database identifies a set of previously unexplored zeolites (i.e., MEI, ATN, EWO, and CAS) that show the highest ensemble-averaged rates for [CuOCu]2+-catalyzed methane activation. We believe that this CBT philosophy can be generally applied to other zeolite-catalyzed reactions and, subsequently, to other types of heterogeneous catalysts. Thus, this represents an important step toward overcoming the long-standing barriers within the computational heterogeneous catalysis community.