Artificial intelligence chemistry最新文献

Automation and high-throughput techniques provide a solid technical foundation for realizing the deep fusion of artificial intelligence and chemistry as well as the full utilization of their advantages. In recent years, with the unique advantages of low consumption, low risk, high efficiency, high reproducibility, high flexibility and good versatility, intelligent automated platforms for high-throughput chemical synthesis aroused widespread concerns of synthetic chemists. In this review, the automated high-throughput chemical synthesis, automated high-throughput sample treatment and characterization technique, as well as the application of artificial intelligence technique in chemical synthesis are introduced. The characteristics of the systems and platforms based on these techniques, including the iChemFoundry platform developed in the ZJU-Hangzhou Global Scientific and Technological Innovation Center, are introduced. The intelligent automated platforms for high-throughput chemical synthesis will reshape the thinking mode of traditional disciplines, promote the innovation of disruptive techniques, redefine the rate of chemical synthesis, and innovate the way of material manufacturing.

The electrocatalytic conversion of CO₂ (CO₂RR) to multi-carbon products has been an appealing strategy to reduce carbon emissions. However, rapid experimental discovery of efficient CO₂RR electrocatalysts and fast recording of full product distribution information is non-trivial. Herein, we used an electrocatalyst testing platform featuring a home-built automatic flow cell to accelerate catalysts screening. Based on 364 effective data points from 42 Cu-lanthanide bimetallic catalysts obtained within 21 working hours, we found that Eu modification over Cu can promote C₂₊ faradaic efficiency (FE). We have previously reported part of the screening data and the optimization of the Mg-Cu catalyst（Angew. Chem. 2022, 134, e202213423）. Here we augmented the dataset by adding the lanthanide modifiers and reported the Eu-Cu catalyst resulted from the high-throughput investigation. Our characterizations revealed that the Eu²⁺ reduced from Eu³⁺ during the catalyst synthesis prevented the agglomeration of nanoparticles, thus making europium modifications stand out from other lanthanide metal modifiers on FE C₂₊ enhancement. We then optimized the Eu-CuO_x catalyst based on the above understanding to achieve ∼80% C₂₊ FE at a high current density of 1.25 A cm⁻².

We have used a Ligand Knowledge Base for bidentate P,P-donor ligands of potential interest to homogeneous catalysis to compare three dimensionality reduction techniques, namely Principal Component Analysis (PCA), Uniform Manifold Approximation and Projection (UMAP) and t-distributed Stochastic Neighbor Embedding (t-SNE). While our previous work on Ligand Knowledge Bases has focused on PCA, here we compare this approach with more recently-published approaches and assess the information retention, visualization, clustering and interpretability which can be achieved for each approach. We find that potential advantages of t-SNE are not realized with a database of the current size (275 entries), and that there is a degree of complementarity between PCA and UMAP. The statistics underlying PCA rely on linear relationships, making interpretation of the resulting plots comparatively straightforward. Since much of chemistry relies on linear structure-property relationships and low-dimensional visualization, the explainability and information retention achieved is attractive. UMAP proved more challenging to interpret, but achieved clear clustering which was often chemically meaningful, and it would be a useful approach for ensuring that distinct subsets of compounds are sampled in a machine-learning context. This analysis also highlighted that the tunability of catalysis achieved through ligand exchange maps well onto some areas of chemical space where closely related ligands cluster, while others represent outliers; these arise from different combinations of steric and electronic effects which chemists will find intuitive.

Designing rubber formulations can greatly benefit from using a database that stores the formulations and corresponding property data of rubber composites. Such a database can expedite the decision-making process by swiftly identifying the most suitable formulations for specific applications. However, the management of a rubber formulation database encounters various issues, including missing formulation and property data, as well as data entry errors. These issues can impede the decision-making processes and even result in incorrect decisions being made. In this study, machine learning (ML) algorithms were applied to analyze rubber formulation databases. Our findings highlight the success of the ML algorithm in effectively filling in missing data and identifying erroneous data. Furthermore, it demonstrates the accurate prediction of properties for untested formulations within the pre-determined database space. The results underline the outstanding performance of ML algorithms in expediting the rubber formulation design process and emphasize their immense potential to play a prominent role in the advancement of rubber composites.

RNA molecules play multifaceted functional and regulatory roles within cells and have garnered significant attention in recent years as promising therapeutic targets. With remarkable successes achieved by artificial intelligence (AI) in different fields such as computer vision and natural language processing, there is a growing imperative to harness AI’s potential in computer-aided drug design (CADD) to discover novel drug compounds that target RNA. Although machine-learning (ML) approaches have been widely adopted in the discovery of small molecules targeting proteins, the application of ML approaches to model interactions between RNA and small molecule is still in its infancy. Compared to protein-targeted drug discovery, the major challenges in ML-based RNA-targeted drug discovery stem from the scarcity of available data resources. With the growing interest and the development of curated databases focusing on interactions between RNA and small molecule, the field anticipates a rapid growth and the opening of a new avenue for disease treatment. In this review, we aim to provide an overview of recent advancements in computationally modeling RNA-small molecule interactions within the context of RNA-targeted drug discovery, with a particular emphasis on methodologies employing ML techniques.

In this work, we investigate the possibility to use an artificial neural network to predict a large number of accurate state-to-state rate constants for atom-diatom collisions, from available rates obtained at two different accuracy levels, using a few accurate rates and many low-accuracy rates. The H + H₂ → H₂ + H chemical reaction is used to benchmark our neural network, as both low and high accuracy state-to-state rates are available in the literature. Our artificial neural network is a multilayer perceptron, using 8 input neurons including the low-accuracy rate constants, with the high accuracy rate constants as the output neuron. The use of machine learning to predict rate constants is very encouraged, as the rates obtained are accurate, even using as low as 1% of the full dataset to train the neural network, and improve greatly the low accuracy rates previously available. This approach can be used to generate full rate constant datasets with a consistent accuracy, from sparse rates obtained with various methods of different accuracies.

Solid-state electrolytes are key ingredients in next-generation devices for energy storage and release. Machine learning molecular dynamics (MLMD) has shown great promise in studying the diffusivity of mobile ions in solid-state electrolytes, with much higher efficiency than conventional ab initio molecular dynamics (AIMD). In this work, we combine an efficient embedded atom neural network (EANN) approach and an uncertainty-driven active learning algorithm that optimally selects data points from high-temperature AIMD trajectories to construct ML potentials for solid-state electrolytes and validate this strategy in a benchmark system, Li₃YCl₆, for which several controversy theoretical results exist. Through systematic MLMD simulations, we find that a typically used small supercell in AIMD simulations fails to predict the supersonic transition at a critical temperature, leading to a significant overestimation of the Li⁺ conductivity in Li₃YCl₆ at room temperature. Fortunately, thanks to the scalability of the EANN potential, extended MLMD simulations in a sufficiently large cell does yield a notable change of temperature-dependence in conductivity at ∼420 K and a much lower room-temperature conductivity in excellent with experiment. Interestingly, our results are all based on a semi-local PBE density functional, which was argued unable to predict the superionic transition. We analyze possible reasons of the seemingly inconsistent MLMD results reported in literature with different ML potentials. This work paves the way of simply using high-temperature AIMD data to generate more reliable MLMD results of low-temperature ionic conductivities in solid-state electrolytes.

Artificial intelligence (AI) is driving a revolution in chemistry, reshaping the landscape of molecular design. This review explores AI’s pivotal roles in the field of organic synthesis applications. AI accurately predicts reaction outcomes, controls chemical selectivity, simplifies synthesis planning, accelerates catalyst discovery, and fuels material innovation and so on. It seamlessly integrates data-driven algorithms with chemical intuition to redefine molecular design. As AI chemistry advances, it promises accelerated research, sustainability, and innovative solutions to chemistry’s pressing challenges. The fusion of AI and chemistry is poised to shape the field’s future profoundly, offering new horizons in precision and efficiency. This review encapsulates the transformation of AI in chemistry, marking a pivotal moment where algorithms and data converge to revolutionize the world of molecules.

Graph neural networks have been successfully applied to machine learning models related to molecules and crystals, due to the similarity between a molecule/crystal and a graph. In this paper, we present three models that are trained with high-quality experimental data to predict three molecular properties (Kováts retention index, normal boiling point, and mass spectrum), using the same GNN architecture. We show that graph representations of molecules, combined with deep learning methodologies and high-quality data sets, lead to accurate machine learning models to predict molecular properties.

Gas adsorption on one-atom-thick membranes is a growing technology for separation applications owing to its excellent energy efficiency. Herein, we investigate the adsorption of the noble gases, Ne, Ar and Kr on graphynes (GYs), a novel class of one-atom-thick carbon membranes using a swarm intelligence technique, namely particle swarm optimization (PSO). Modeling the adsorption of noble gas clusters on two-dimensional substrates requires a thorough examination of the energy landscape. The high dimensionality of the problem makes it tricky to employ ab initio methods for such studies, necessitating the use of a metaheuristic global optimization technique such as PSO. We explored the adsorption of 1–30 atoms of Ne, Ar and Kr on α-, β-, γ- and rhombic-GYs to predict the most suitable form of GY for the adsorption of each of the gases. Employing the dispersion-corrected density functional theory (DFT-D) data for the adsorption of single gas atoms as the reference data, we parametrized two empirical pairwise potentials, namely, Lennard-Jones (LJ) and improved Lennard-Jones (ILJ) potentials. We then analyzed the growth pattern as well as the energetics of adsorption using the parametrized potentials, in combination with the PSO technique, which enabled us to predict the best possible membrane for the adsorption of the noble gases: α-GY for Ne and γ-GY for Ar and Kr. The accuracy of our modeling approach is further validated against DFT-D computations thereby establishing that PSO, when combined with the ILJ potential, can serve as a computationally feasible approach for modeling gas adsorption on GYs.