Many-Body Function Corrected Neural Network with Atomic Attention (MBNN-att) for Molecular Property Prediction.

Recent years have seen a surge of machine learning (ML) in chemistry for predicting chemical properties, but a low-cost, general-purpose, and high-performance model, desirable to be accessible on central processing unit (CPU) devices, remains not available. For this purpose, here we introduce an atomic attention mechanism into many-body function corrected neural network (MBNN), namely, MBNN-att ML model, to predict both the extensive and intensive properties of molecules and materials. The MBNN-att uses explicit function descriptors as the inputs for the atom-based feed-forward neural network (NN). The output of the NN is designed to be a vector to implement the multihead self-attention mechanism. This vector is split into two parts: the atomic attention weight part and the many-body-function part. The final property is obtained by summing the products of each atomic attention weight and the corresponding many-body function. We show that MBNN-att performs well on all QM9 properties, i.e., errors on all properties, below chemical accuracy, and, in particular, achieves the top performance for the energy-related extensive properties. By systematically comparing with other explicit-function-type descriptor ML models and the graph representation ML models, we demonstrate that the many-body-function framework and atomic attention mechanism are key ingredients for the high performance and the good transferability of MBNN-att in molecular property prediction.

Global Neural Network Potential with Explicit Many-Body Functions for Improved Descriptions of Complex Potential Energy Surface.

The high dimensional machine learning potential (MLP) that has developed rapidly in the past decade represents a giant step forward in large-scale atomic simulation for complex systems. The long-range interaction and the poor description of chemical reactions are typical problems of high dimensional MLP, which are mainly caused by the poor structure discrimination of the atom-centered ML model. Herein, we propose a low-cost neural-network-based MLP architecture for fitting global potential energy surface data, namely, G-MBNN, that can offer improved energy and force resolution on a complex potential energy surface. In G-MBNN, a set of many-body energy terms based on the local atomic environment are explicitly included in computing the total energy─the total energy of the system is written as the sum of atomic energy and many-body energy contributions. These extra many-body energy terms are computationally low-cost and, importantly, can provide easy access to delicate energy terms in complex systems such as very short repulsion, long-range attractions, and sensitive angular-dependent covalent interactions. We implement G-MBNN in the LASP code and demonstrate the improved accuracy of the new framework in representative systems, including ternary-element energy materials LiCoOx, TiO2 with defects, and a series of organic reactions.

Artificial intelligence pathway search to resolve catalytic glycerol hydrogenolysis selectivity.

The complex interaction between molecules and catalyst surfaces leads to great difficulties in understanding and predicting the activity and selectivity in heterogeneous catalysis. Here we develop an end-to-end artificial intelligence framework for the activity prediction of heterogeneous catalytic systems (AI-Cat method), which takes simple inputs from names of molecules and metal catalysts and outputs the reaction energy profile from the input molecule to low energy pathway products. The AI-Cat method combines two neural network models, one for predicting reaction patterns and the other for providing the reaction barrier and energy, with a Monte Carlo tree search to resolve the low energy pathways in a reaction network. We then apply AI-Cat to resolve the reaction network of glycerol hydrogenolysis on Cu surfaces, which is a typical selective C-O bond activation system and of key significance for biomass-derived polyol utilization. We show that glycerol hydrogenolysis features a huge reaction network of relevant candidates, containing 420 reaction intermediates and 2467 elementary reactions. Among them, the surface-mediated enol-keto tautomeric resonance is a key step to facilitate the primary C-OH bond breaking and thus selects 1,2-propanediol as the major product on Cu catalysts. 1,3-Propanediol can only be produced under strong acidic conditions and high surface H coverage by following a hydrogenation-dehydration pathway. AI-Cat further discovers six low-energy reaction patterns for C-O bond activation on metals that is of general significance to polyol catalysis. Our results demonstrate that the reaction prediction for complex heterogeneous catalysis is now feasible with AI-based atomic simulation and a Monte Carlo tree search.

Methanol Synthesis from CO2/CO Mixture on Cu-Zn Catalysts from Microkinetics-Guided Machine Learning Pathway Search.

Methanol synthesis on industrial Cu/ZnO/Al2O3 catalysts via the hydrogenation of CO and CO2 mixture, despite several decades of research, is still puzzling due to the nature of the active site and the role of CO2 in the feed gas. Herein, with the large-scale machine learning atomic simulation, we develop a microkinetics-guided machine learning pathway search to explore thousands of reaction pathways for CO2 and CO hydrogenations on thermodynamically favorable Cu-Zn surface structures, including Cu(111), Cu(211), and Zn-alloyed Cu(211) surfaces, from which the lowest energy pathways are identified. We find that Zn decorates at the step-edge at Cu(211) up to 0.22 ML under reaction conditions with the Zn-Zn dimeric sites being avoided. CO2 and CO hydrogenations occur exclusively at the step-edge of the (211) surface with up to 0.11 ML Zn coverage, where the low coverage of Zn (0.11 ML) does not much affect the reaction kinetics, but the higher coverages of Zn (0.22 ML) poison the catalyst. It is CO2 hydrogenation instead of CO hydrogenation that dominates methanol synthesis, agreeing with previous isotope experiments. While metallic steps are identified as the major active site, we show that the [-Zn-OH-Zn-] chains (cationic Zn) can grow on Cu(111) surfaces under reaction conditions, which suggests the critical role of CO in the mixed gas for reducing the cationic Zn and exposing metal sites for methanol synthesis. Our results provide a comprehensive picture on the dynamic coupling of the feed gas composition, the catalyst active site, and the reaction activity in this complex heterogeneous catalytic system.

Reaction prediction via atomistic simulation: from quantum mechanics to machine learning.

It is an ultimate goal in chemistry to predict reaction without recourse to experiment. Reaction prediction is not just the reaction rate determination of known reactions but, more broadly, the reaction exploration to identify new reaction routes. This review briefly overviews the theory on chemical reaction and the current methods for computing/estimating reaction rate and exploring reaction space. We particularly focus on the atomistic simulation methods for reaction exploration, which are benefited significantly by recently emerged machine learning potentials. We elaborate the stochastic surface walking global pathway sampling based on the global neural network (SSW-NN) potential, developed in our group since 2013, which can explore complex reactions systems unbiasedly and automatedly. Two examples, molecular reaction and heterogeneous catalytic reactions, are presented to illustrate the current status for reaction prediction using SSW-NN.

Large-Scale Atomic Simulation via Machine Learning Potentials Constructed by Global Potential Energy Surface Exploration.

Atomic simulations based on quantum mechanics (QM) calculations have entered into the tool box of chemists over the past few decades, facilitating an understanding of a wide range of chemistry problems, from structure characterization to reactivity determination. Due to the poor scaling and high computational cost intrinsic to QM calculations, one has to either sacrifice accuracy or time when performing large-scale atomic simulations. The battle to find a better compromise between accuracy and speed has been central to the development of new theoretical methods.The recent advances of machine-learning (ML)-based large-scale atomic simulations has shown great promise to the benefit of many branches of chemistry. Instead of solving the Schrödinger equation directly, ML-based simulations rely on a large data set of accurate potential energy surfaces (PESs) and complex numerical models to predict the total energy. These simulations feature both a high speed and a high accuracy for computing large systems. Due to the lack of a physical foundation in numerical models, ML models are often frustrated in their predictivity and robustness, which are key to applications. Focusing on these concerns, here we overview the recent advances in ML methodologies for atomic simulations on three key aspects. Namely, the generation of a representative data set, the extensity of ML models, and the continuity of data representation. While global optimization methods are the natural choice for building a representative data set, the stochastic surface walking method is shown to provide the desired PES sampling for both minima and transition regions on the PES. The current ML models generally utilize local geometrical descriptors as an input and consider the total energy as the sum of atomic energies. There are many flavors of data descriptors and ML models, but the applications for material and reaction predictions are still limited, not least because of the difficulty to train the associated vast global data sets. We show that our recently designed power-type structure descriptors together with a feed-forward neural network (NN) model are compatible with highly complex global PES data, which has led to a large family of global NN (G-NN) potentials.Two recent applications of G-NN potentials in material and reaction simulations are selected to illustrate how ML-based atomic simulations can help the discovery of new materials and reactions.

Glucose to 5-Hydroxymethylfurfural: Origin of Site-Selectivity Resolved by Machine Learning Based Reaction Sampling.

Glucose pyrolysis, a model system in biomass utilization, is renowned for its great complexity, deep in reaction network hierarchy and rich in reaction patterns. The selectivity in glucose pyrolysis, e.g., the high yield of 5-hydroxymethylfurfural (HMF), a value-added platform product, remains an intriguing puzzle even after 60 years of experimental study. Here we resolve the whole reaction network of glucose pyrolysis using a global-to-global technique for reaction pathway sampling. This is achieved by establishing the first organic chemistry reaction database via stochastic surface walking (SSW) global optimization, building the global neural network (G-NN) potential via machine learning and extensively exploring the reaction network of glucose pyrolysis. In total, 6407 elementary reactions, screened out from more than 150 000 reaction pairs in glucose pyrolysis, are collected in our reaction database. The established reaction network from SSW-NN, further validated by first-principles calculations, reveals that for glucose to HMF, the lowest energy reaction pathway involves fructose and 3-deoxyglucos-2-ene (3-DGE) as key intermediates and a site-selective reaction type, retro-Michael-addition, for three consecutive dehydration steps. The overall barrier is determined to be 1.91 eV, being at least 0.19 eV lower than all previously proposed mechanisms, which assumes direct ß-H elimination dehydration. The lowest pathways to the other two major products, furfural (FF) and hydroxyacetaldehyde (HAA), are also discovered with a similar barrier 1.95 eV, which exhibit a competing nature by sharing the same key intermediate, 3-ketohexose. Since chemical reactions occurring in fast glucose pyrolysis are generally present in biomass chemistry, containing essentially all reaction patterns of C-H-O elements, the methodology designed and the results presented would help to advance reaction design and mechanistic modeling in renewable fuels from biomass.

Atomic structure of boron resolved using machine learning and global sampling.

Boron crystals, despite their simple composition, must rank top for complexity: even the atomic structure of the ground state of ß-B remains uncertain after 60 years' study. This makes it difficult to understand the many exotic photoelectric properties of boron. The presence of self-doping atoms in the crystal interstitial sites forms an astronomical configurational space, making the determination of the real configuration virtually impossible using current techniques. Here, by combining machine learning with the latest stochastic surface walking (SSW) global optimization, we explore for the first time the potential energy surface of ß-B, revealing 15 293 distinct configurations out of the 2 × 105 minima visited, and reveal the key rules governing the filling of the interstitial sites. This advance is only allowed by the construction of an accurate and efficient neural network (NN) potential using a new series of structural descriptors that can sensitively discriminate the complex boron bonding environment. We show that, in contrast to the conventional views on the numerous energy-degenerate configurations, only 40 minima of ß-B are identified to be within 7 meV per atom in energy above the global minimum of ß-B, most of them having been discovered for the first time. These low energy structures are classified into three types of skeletons and six patterns of doping configurations, with a clear preference for a few characteristic interstitial sites. The observed ß-B and its properties are influenced strongly by a particular doping site, the B19 site that neighbors the B18 site, which has an exceptionally large vibrational entropy. The configuration with this B19 occupancy, which ranks only 15th at 0 K, turns out to be dominant at high temperatures. Our results highlight the novel SSW-NN architecture as the leading problem solver for complex material phenomena, which would then expedite substantially the building of a material genome database.