Reaction Templates: Bridging Synthesis Knowledge and Artificial Intelligence.

ConspectusThe field of chemical research boasts a long history of developing software to automate synthesis planning and reaction prediction. Early software relied heavily on expert systems, requiring significant effort to encode vast amounts of synthesis knowledge into a computer-readable format. However, recent advancements in deep learning have shifted the focus toward AI models, offering improved prediction capabilities. Despite these advancements, current AI models often lack the integration of known synthesis rules and intuitions, creating a gap that hinders interpretability and future development of the models. To bridge them, our research group has been actively working on incorporating reaction templates into deep learning models, achieving promising results across various applications.In this Account, we present our latest works to incorporate the known synthesis knowledge into the deep learning models through the utilization of reaction templates. We begin by highlighting the limitations of early computer programs heavily reliant on hand-coded rules. These programs, while providing a foundation for the field, presented limitations in scalability and adaptability. We then introduce SMARTS (SMILES arbitrary target specification), a popular Python-readable format for representing chemical reactions. This format of reaction encoding facilitates the quick integration of synthesis knowledge into AI models built using the Python language. With the SMARTS-based reaction templates, we introduce our recent efforts of developing an AI model for reaction-based molecule optimization. Subsequently, we discuss the recent efforts to automate the extraction of reaction templates from vast chemical reaction databases. This approach eliminates the previously required manual effort of encoding knowledge, a process that could be time-consuming and prone to error when dealing with large data sets. By customizing the automated extraction algorithm, we have developed powerful AI models for specific tasks such as retrosynthesis (LocalRetro), reaction outcome prediction (LocalTransform), and atom-to-atom mapping (LocalMapper). These models, aligned with the intuition of chemists, demonstrate the effectiveness of incorporating reaction templates into deep learning frameworks.Looking toward the future, we believe that utilizing reaction templates to connect known chemical knowledge and AI models holds immense potential for various applications. Not only can this approach significantly benefit future AI models focused on challenging tasks like reaction mechanism labeling and prediction, but we anticipate it can also extend its reach to the realm of inorganic synthesis. By integrating synthesis knowledge, we can not only achieve improved performance but also enhance the interpretability of AI models, paving the way for further advancements in AI-powered chemical synthesis.

Surface Area-Enhanced Cerium and Sulfur-Modified Hierarchical Bismuth Oxide Nanosheets for Electrochemical Carbon Dioxide Reduction to Formate.

Electrochemical carbon dioxide reduction reaction (ECO2RR) is a promising approach to synthesize fuels and value-added chemical feedstocks while reducing atmospheric CO2 levels. Here, high surface area cerium and sulfur-doped hierarchical bismuth oxide nanosheets (Ce@S-Bi2O3) are develpoed by a solvothermal method. The resulting Ce@S-Bi2O3 electrocatalyst shows a maximum formate Faradaic efficiency (FE) of 92.5% and a current density of 42.09 mA cm-2 at -1.16 V versus RHE using a traditional H-cell system. Furthermore, using a three-chamber gas diffusion electrode (GDE) reactor, a maximum formate FE of 85% is achieved in a wide range of applied potentials (-0.86 to -1.36 V vs RHE) using Ce@S-Bi2O3. The density functional theory (DFT) results show that doping of Ce and S in Bi2O3 enhances formate production by weakening the OH* and H* species. Moreover, DFT calculations reveal that *OCHO is a dominant pathway on Ce@S-Bi2O3 that leads to efficient formate production. This study opens up new avenues for designing metal and element-doped electrocatalysts to improve the catalytic activity and selectivity for ECO2RR.

Autobifunctional Mechanism of Jagged Pt Nanowires for Hydrogen Evolution Kinetics via End-to-End Simulation.

The extraordinary mass activity of jagged Pt nanowires can substantially improve the economics of the hydrogen evolution reaction (HER). However, it is a great challenge to fully unveil the HER kinetics driven by the jagged Pt nanowires with their multiscale morphology. Herein we present an end-to-end framework that combines experiment, machine learning, and multiscale advances of the past decade to elucidate the HER kinetics catalyzed by jagged Pt nanowires under alkaline conditions. The bifunctional catalysis conventionally refers to the synergistic increase in activity by the combination of two different catalysts. We report that monometals, such as jagged Pt nanowires, can exhibit bifunctional characteristics owing to its complex surface morphology, where one site prefers electrochemical proton adsorption and another is responsible for activation, resulting in a 4-fold increase in the activity. We find that the conventional design guideline that the sites with a 0 eV Gibbs free energy of adsorption are optimal for HER does not hold under alkaline conditions, and rather, an energy between -0.2 and 0.0 eV is shown to be optimal. At the reaction temperatures, the high activity arises from low-coordination-number (≤7) Pt atoms exposed by the jagged surface. Our current demonstration raises an emerging prospect to understand highly complex kinetic phenomena on the nanoscale in full by implementing end-to-end multiscale strategies.

Structure-Based Synthesizability Prediction of Crystals Using Partially Supervised Learning.

Predicting the synthesizability of inorganic materials is one of the major challenges in accelerated material discovery. A widely employed approximate approach is to consider the thermodynamic decomposition stability due to its simplicity of computing, but it is notorious for either producing too many candidates or missing important metastable materials. These results, however, are not unexcepted since the synthesizability is a complex phenomenon, and the thermodynamic stability is just one contributor. Here, we suggest a machine-learning model to quantify the probability of synthesis based on the partially supervised learning of materials database. We adapted the positive and unlabeled machine learning (PU learning) by implementing the graph convolutional neural network as a classifier in which the model outputs crystal-likeness scores (CLscore). The model shows 87.4% true positive (CLscore > 0.5) prediction accuracy for the test set of experimentally reported cases (9356 materials) in the Materials Project. We further validated the model by predicting the synthesizability of newly reported experimental materials in the last 5 years (2015-2019) with an 86.2% true positive rate using the model trained with the database as of the end of year 2014. Our analysis shows that our model captures the structural motif for synthesizability beyond what is possible by Ehull. We find that 71 materials among the top 100 high-scoring virtual materials have indeed been previously synthesized in the literature. With the proposed data-driven metric of the crystal-likeness score, high-throughput virtual screenings and generative models can benefit significantly by effectively reducing the chemical space that needs to be explored experimentally in the future toward more rational materials design.

Uncertainty-Quantified Hybrid Machine Learning/Density Functional Theory High Throughput Screening Method for Crystals.

Computational high throughput screening (HTS) has emerged as a significant tool in material science to accelerate the discovery of new materials with target properties in recent years. However, despite many successful cases in which HTS led to the novel discovery, currently, the major bottleneck in HTS is a large computational cost of density functional theory (DFT) calculations that scale cubically with system size, limiting the chemical space that can be explored. The present work aims at addressing this computational burden of HTS by presenting a machine learning (ML) framework that can efficiently explore the chemical space. Our model is built upon an existing crystal graph convolutional neural network (CGCNN) to obtain formation energy of a crystal structure but is modified to allow uncertainty quantification for each prediction using the hyperbolic tangent activation function and dropout algorithm (CGCNN-HD). The uncertainty quantification is particularly important since typical usage of CGCNN (due to the lack of gradient implementation) does not involve structural relaxation which could cause substantial prediction errors. The proposed method is benchmarked against an existing application that identified promising photoanode material among the >7,000 hypothetical Mg-Mn-O ternary compounds using all DFT-HTS. In our approach, we perform the approximate HTS using CGCNN-HD and refine the results using full DFT for those selected (denoted as ML/DFT-HTS). The proposed hybrid model reduces the required DFT calculations by a factor of >50 compared to the previous DFT-HTS in making the same discovery of Mg2MnO4, experimentally validated new photoanode material. Further analysis demonstrates that the addition of HD components with uncertainty measures in the CGCNN-HD model increased the discoverability of promising materials relative to all DFT-HTS from 30% (CGCNN) to 68% (CGCNN-HD). The present ML/DFT-HTS with uncertainty quantification can thus be a fast alternative to DFT-HTS for efficient exploration of the vast chemical space.

Predicting synthesis recipes of inorganic crystal materials using elementwise template formulation.

While advances in computational techniques have accelerated virtual materials design, the actual synthesis of predicted candidate materials is still an expensive and slow process. While a few initial studies attempted to predict the synthesis routes for inorganic crystals, the existing models do not yield the priority of predictions and could produce thermodynamically unrealistic precursor chemicals. Here, we propose an element-wise graph neural network to predict inorganic synthesis recipes. The trained model outperforms the popularity-based statistical baseline model for the top-k exact match accuracy test, showing the validity of our approach for inorganic solid-state synthesis. We further validate our model by the publication-year-split test, where the model trained based on the materials data until the year 2016 is shown to successfully predict synthetic precursors for the materials synthesized after 2016. The high correlation between the probability score and prediction accuracy suggests that the probability score can be interpreted as a measure of confidence levels, which can offer the priority of the predictions.

Accelerated chemical science with AI.

In light of the pressing need for practical materials and molecular solutions to renewable energy and health problems, to name just two examples, one wonders how to accelerate research and development in the chemical sciences, so as to address the time it takes to bring materials from initial discovery to commercialization. Artificial intelligence (AI)-based techniques, in particular, are having a transformative and accelerating impact on many if not most, technological domains. To shed light on these questions, the authors and participants gathered in person for the ASLLA Symposium on the theme of 'Accelerated Chemical Science with AI' at Gangneung, Republic of Korea. We present the findings, ideas, comments, and often contentious opinions expressed during four panel discussions related to the respective general topics: 'Data', 'New applications', 'Machine learning algorithms', and 'Education'. All discussions were recorded, transcribed into text using Open AI's Whisper, and summarized using LG AI Research's EXAONE LLM, followed by revision by all authors. For the broader benefit of current researchers, educators in higher education, and academic bodies such as associations, publishers, librarians, and companies, we provide chemistry-specific recommendations and summarize the resulting conclusions.

Automated exploitation of the big configuration space of large adsorbates on transition metals reveals chemistry feasibility.

Mechanistic understanding of large molecule conversion and the discovery of suitable heterogeneous catalysts have been lagging due to the combinatorial inventory of intermediates and the inability of humans to enumerate all structures. Here, we introduce an automated framework to predict stable configurations on transition metal surfaces and demonstrate its validity for adsorbates with up to 6 carbon and oxygen atoms on 11 metals, enabling the exploration of ~108 potential configurations. It combines a graph enumeration platform, force field, multi-fidelity DFT calculations, and first-principles trained machine learning. Clusters in the data reveal groups of catalysts stabilizing different structures and expose selective catalysts for showcase transformations, such as the ethylene epoxidation on Ag and Cu and the lack of C-C scission chemistry on Au. Deviations from the commonly assumed atom valency rule of small adsorbates are also manifested. This library can be leveraged to identify catalysts for converting large molecules computationally.

Correction to "Generative Adversarial Networks for Crystal Structure Prediction".

[This corrects the article DOI: 10.1021/acscentsci.0c00426.].

Bimetallic Gold-Silver Nanostructures Drive Low Overpotentials for Electrochemical Carbon Dioxide Reduction.

Alloy formation is an advanced approach to improve desired properties that the monoelements cannot achieve. Alloys are usually designed to tailor intrinsic natures or induce synergistic effects by combining materials with distinct properties. Indeed, unprecedented properties have emerged in many cases, superior to a simple sum of pure elements. Here, we present Au-Ag alloy nanostructures with prominent catalytic properties in an electrochemical carbon dioxide reduction reaction (eCO2RR). The Au-Ag hollow nanocubes are prepared by galvanic replacement of Au on Ag nanocubes. When the Au-to-Ag ratio is 1:1 (Au1Ag1), the alloy hollow nanocubes exhibit maximum Faradaic efficiencies of CO production in a wide potential range and high mass activity and CO current density superior to those of the bare metals. In particular, overpotentials are estimated to be similar to or lower than that of the Au catalyst under various standard metrics. Density functional theory calculations, machine learning, and a statistical consideration demonstrate that the optimal configuration of the *COOH intermediate is a bidentate coordination structure where C binds to Au and O binds to Ag. This active Au-Ag neighboring configuration has a maximum population and enhanced intrinsic catalytic activity on the Au1Ag1 surface among other Au-to-Ag compositions, in good agreement with the experimental results. Further application of Au1Ag1 to a membrane electrode assembly cell at neutral conditions shows enhanced CO Faradaic efficiency and current densities compared to Au or Ag nanocubes, indicating the possible extension of Au-Ag alloys to larger electrochemical systems. These results give a new insight into the synergistic roles of Au and Ag in the eCO2RR and offer a fresh direction toward a rational design of bimetallic catalysts at a practical scale.

Understanding potential-dependent competition between electrocatalytic dinitrogen and proton reduction reactions.

A key challenge to realizing practical electrochemical N2 reduction reaction (NRR) is the decrease in the NRR activity before reaching the mass-transfer limit as overpotential increases. While the hydrogen evolution reaction (HER) has been suggested to be responsible for this phenomenon, the mechanistic origin has not been clearly explained. Herein, we investigate the potential-dependent competition between NRR and HER using the constant electrode potential model and microkinetic modeling. We find that the H coverage and N2 coverage crossover leads to the premature decrease of NRR activity. The coverage crossover originates from the larger charge transfer in H+ adsorption than N2 adsorption. The larger charge transfer in H+ adsorption, which potentially leads to the coverage crossover, is a general phenomenon seen in various heterogeneous catalysts, posing a fundamental challenge to realize practical electrochemical NRR. We suggest several strategies to overcome the challenge based on the present understandings.

Predicting potentially hazardous chemical reactions using an explainable neural network.

Predicting potentially dangerous chemical reactions is a critical task for laboratory safety. However, a traditional experimental investigation of reaction conditions for possible hazardous or explosive byproducts entails substantial time and cost, for which machine learning prediction could accelerate the process and help detailed experimental investigations. Several machine learning models have been developed which allow the prediction of major chemical reaction products with reasonable accuracy. However, these methods may not present sufficiently high accuracy for the prediction of hazardous products which particularly requires a low false negative result for laboratory safety in order not to miss any dangerous reactions. In this work, we propose an explainable artificial intelligence model that can predict the formation of hazardous reaction products in a binary classification fashion. The reactant molecules are transformed into substructure-encoded fingerprints and then fed into a convolutional neural network to make the binary decision of the chemical reaction. The proposed model shows a false negative rate of 0.09, which can be compared with 0.47-0.66 using the existing main product prediction models. To provide explanations for what substructures of the given reactant molecules are important to make a decision for target hazardous product formation, we apply an input attribution method, layer-wise relevance propagation, which computes the contributions of individual inputs per input data. The computed attributions indeed match some of the existing chemical intuitions and mechanisms, and also offer a way to analyze possible data-imbalance issues of the current predictions based on relatively small positive datasets. We expect that the proposed hazardous product prediction model will be complementary to existing main product prediction models and experimental investigations.

Machine-enabled inverse design of inorganic solid materials: promises and challenges.

Developing high-performance advanced materials requires a deeper insight and search into the chemical space. Until recently, exploration of materials space using chemical intuitions built upon existing materials has been the general strategy, but this direct design approach is often time and resource consuming and poses a significant bottleneck to solve the materials challenges of future sustainability in a timely manner. To accelerate this conventional design process, inverse design, which outputs materials with pre-defined target properties, has emerged as a significant materials informatics platform in recent years by leveraging hidden knowledge obtained from materials data. Here, we summarize the latest progress in machine-enabled inverse materials design categorized into three strategies: high-throughput virtual screening, global optimization, and generative models. We analyze challenges for each approach and discuss gaps to be bridged for further accelerated and rational data-driven materials design.

Generative Adversarial Networks for Crystal Structure Prediction.

The constant demand for novel functional materials calls for efficient strategies to accelerate the materials discovery, and crystal structure prediction is one of the most fundamental tasks along that direction. In addressing this challenge, generative models can offer new opportunities since they allow for the continuous navigation of chemical space via latent spaces. In this work, we employ a crystal representation that is inversion-free based on unit cell and fractional atomic coordinates and build a generative adversarial network for crystal structures. The proposed model is applied to generate the Mg-Mn-O ternary materials with the theoretical evaluation of their photoanode properties for high-throughput virtual screening (HTVS). The proposed generative HTVS framework predicts 23 new crystal structures with reasonable calculated stability and band gap. These findings suggest that the generative model can be an effective way to explore hidden portions of the chemical space, an area that is usually unreachable when conventional substitution-based discovery is employed.

Practical Deep-Learning Representation for Fast Heterogeneous Catalyst Screening.

The binding site and energy is an invaluable descriptor in high-throughput screening of catalysts, as it is accessible and correlates with the activity and selectivity. Recently, comprehensive binding energy prediction machine-learning models have been demonstrated and promise to accelerate the catalyst screening. Here, we present a simple and versatile representation, applicable to any deep-learning models, to further accelerate such process. Our approach involves labeling the binding site atoms of the unrelaxed bare surface geometry; hence, for the model application, density functional theory calculations can be completely removed if the optimized bulk structure is available as is the case when using the Materials Project database. In addition, we present ensemble learning, where a set of predictions is used together to form a predictive distribution that reduces the model bias. We apply the labeled site approach and ensemble to crystal graph convolutional neural network and the ∼40 000 data set of alloy catalysts for CO2 reduction. The proposed model applied to the data set of unrelaxed structures shows 0.116 and 0.085 eV mean absolute error, respectively, for CO and H binding energy, better than the best method (0.13 and 0.13 eV) in the literature that requires costly geometry relaxations. The analysis of the model parameters demonstrates that the model can effectively learn the chemical information related to the binding site.

Progress in Computational and Machine-Learning Methods for Heterogeneous Small-Molecule Activation.

The chemical conversion of small molecules such as H2 , H2 O, O2 , N2 , CO2 , and CH4 to energy and chemicals is critical for a sustainable energy future. However, the high chemical stability of these molecules poses grand challenges to the practical implementation of these processes. In this regard, computational approaches such as density functional theory, microkinetic modeling, data science, and machine learning have guided the rational design of catalysts by elucidating mechanistic insights, identifying active sites, and predicting catalytic activity. Here, the theory and methodologies for heterogeneous catalysis and their applications for small-molecule activation are reviewed. An overview of fundamental theory and key computational methods for designing catalysts, including the emerging data science techniques in particular, is given. Applications of these methods for finding efficient heterogeneous catalysts for the activation of the aforementioned small molecules are then surveyed. Finally, promising directions of the computational catalysis field for further outlooks are discussed, focusing on the challenges and opportunities for new methods.

Unveiling new stable manganese based photoanode materials via theoretical high-throughput screening and experiments.

With the increasing energy demand, developing renewable fuel production strategies such as photoelectrocatalytic hydrogen production is critical to mitigating the global climate change. In this work, we experimentally validate a new stable and photoactive material, Mg2MnO4, from the exhaustive theoretical exploration of the chemical space of X (=Mg and Ca), Mn and O.

Guaiacol hydrodeoxygenation mechanism on Pt(111): insights from density functional theory and linear free energy relations.

Density functional theory is used to study the adsorption of guaiacol and its initial hydrodeoxygenation (HDO) reactions on Pt(111). Previous Brønsted-Evans-Polanyi (BEP) correlations for small open-chain molecules are inadequate in estimating the reaction barriers of phenolic compounds except for the side group (methoxy) carbon-dehydrogenation. New BEP relations are established using a select group of phenolic compounds. These relations are applied to construct a potential-energy surface of guaiacol-HDO to catechol. Analysis shows that catechol is mainly produced via dehydrogenation of the methoxy functional group followed by the CHx (x<3) removal of the functional group and hydrogenation of the ring carbon, in contrast to a hypothesis of a direct demethylation path. Dehydroxylation and demethoxylation are slow, implying that phenol is likely produced from catechol but not through its direct dehydroxylation followed by aromatic carbon-ring hydrogenation.