Digital discovery最新文献

Forecasting experimental chemical shifts of organic compounds is a long-standing challenge in organic chemistry. Recent advances in machine learning (ML) have led to routines that surpass the accuracy of ab initio Density Functional Theory (DFT) in estimating experimental ¹³C shifts. The extraction of knowledge from other models, known as transfer learning, has demonstrated remarkable improvements, particularly in scenarios with limited data availability. However, the extent to which transfer learning improves predictive accuracy in low-data regimes for experimental chemical shift predictions remains unexplored. This study indicates that atomic features derived from a message passing neural network (MPNN) forcefield are robust descriptors for atomic properties. A dense network utilizing these descriptors to predict ¹³C shifts achieves a mean absolute error (MAE) of 1.68 ppm. When these features are used as node labels in a simple graph neural network (GNN), the model attains a better MAE of 1.34 ppm. On the other hand, embeddings from a self-supervised pre-trained 3D aware transformer are not sufficiently descriptive for a feedforward model but show reasonable accuracy within the GNN framework, achieving an MAE of 1.51 ppm. Under low-data conditions, all transfer-learned models show a significant improvement in predictive accuracy compared to existing literature models, regardless of the sampling strategy used to select from the pool of unlabeled examples. We demonstrated that extracting atomic features from models trained on large and diverse datasets is an effective transfer learning strategy for predicting NMR chemical shifts, achieving results on par with existing literature models. This method provides several benefits, such as reduced training times, simpler models with fewer trainable parameters, and strong performance in low-data scenarios, without the need for costly ab initio data of the target property. This technique can be applied to other chemical tasks opening many new potential applications where the amount of data is a limiting factor.

The investigation of inhomogeneous surfaces, where various local structures coexist, is crucial for understanding interfaces of technological interest, yet it presents significant challenges. Here, we study the atomic configurations of the (2 × m) Ti-rich surfaces at (110)-oriented SrTiO₃ by bringing together scanning tunneling microscopy and transferable neural-network force fields combined with evolutionary exploration. We leverage an active learning methodology to iteratively extend the training data as needed for different configurations. Training on only small well-known reconstructions, we are able to extrapolate to the complicated and diverse overlayers encountered in different regions of the inhomogeneous SrTiO₃(110)-(2 × m) surface. Our machine-learning-backed approach generates several new candidate structures, in good agreement with experiment and verified using density functional theory. The approach could be extended to other complex metal oxides featuring large coexisting surface reconstructions.

We demonstrate a novel mechanochemical synthesis method using a robotic powder grinding system that applies a precisely controlled and constant mechanical force. This approach significantly enhances reproducibility and enables detailed analysis of reaction pathways. Our results indicate that robotic force control can alter the reaction rate and influence the reaction pathway, highlighting its potential for elucidating chemical reaction mechanisms and fostering the discovery of new chemical reactions. Despite its significance, the application of a controllable constant force in macroscale mechanochemical synthesis remains challenging. To address this gap, we compared the reproducibilities of various mechanochemical syntheses using conventional manual grinding, ball milling, and our novel robotic approach with perovskite materials. Our findings indicate that the robotic approach provides significantly higher reproducibility than conventional methods, facilitating the analysis of reaction pathways. By manipulating the grinding force and speed, we revealed that robotic force control can alter both the reaction rate and pathway. Consequently, robotic mechanochemical synthesis has significant potential for advancing the understanding of chemical reaction mechanisms and discovering new reactions.

Zeolites are complex and porous crystalline inorganic materials that serve as hosts for a variety of molecular, ionic and cluster species. Formal, machine-actionable representation of this chemistry presents a challenge as a variety of concepts need to be semantically interlinked. This work demonstrates the potential of knowledge engineering in overcoming this challenge. We develop ontologies OntoCrystal and OntoZeolite, enabling the representation and instantiation of crystalline zeolite information into a dynamic, interoperable knowledge graph called The World Avatar (TWA). In TWA, crystalline zeolite instances are semantically interconnected with chemical species that act as guests in these materials. Information can be obtained via custom or templated SPARQL queries administered through a user-friendly web interface. Unstructured exploration is facilitated through natural language processing using the Marie System, showcasing promise for the blended large language model – knowledge graph approach in providing accurate responses on zeolite chemistry in natural language.

The ability and desire to collect kinetic data has greatly increased in recent years, requiring more automated and quantitative methods for analysis. In this work, an automated program (Auto-VTNA) is developed, to simplify the kinetic analysis workflow. Auto-VTNA allows all the reaction orders to be determined concurrently, expediting the process of kinetic analysis. Auto-VTNA performs well on noisy or sparse data sets and can handle complex reactions involving multiple reaction orders. Quantitative error analysis and facile visualisation allows users to numerically justify and robustly present their findings. Auto-VTNA can be used through a free graphical user interface (GUI), requiring no coding or expert kinetic model input from the user, and can be customised and built on if required.

The rise of electron microscopy has expanded our ability to acquire nanometer and atomically resolved images of complex materials. The resulting vast datasets are typically analyzed by human operators, an intrinsically challenging process due to the multiple possible analysis steps and the corresponding need to build and optimize complex analysis workflows. We present a methodology based on the concept of a Reward Function coupled with Bayesian Optimization, to optimize image analysis workflows dynamically. The Reward Function is engineered to closely align with the experimental objectives and broader context and is quantifiable upon completion of the analysis. Here, cross-section, high-angle annular dark field (HAADF) images of ion-irradiated (Y, Dy)Ba₂Cu₃O_7−δ thin-films were used as a model system. The reward functions were formed based on the expected materials density and atomic spacings and used to drive multi-objective optimization of the classical Laplacian-of-Gaussian (LoG) method. These results can be benchmarked against the DCNN segmentation. This optimized LoG* compares favorably against DCNN in the presence of the additional noise. We further extend the reward function approach towards the identification of partially-disordered regions, creating a physics-driven reward function and action space of high-dimensional clustering. We pose that with correct definition, the reward function approach allows real-time optimization of complex analysis workflows at much higher speeds and lower computational costs than classical DCNN-based inference, ensuring the attainment of results that are both precise and aligned with the human-defined objectives.

The effectiveness of drug treatments depends significantly on the water solubility of compounds, influencing bioavailability and therapeutic outcomes. A reliable predictive solubility tool enables drug developers to swiftly identify drugs with low solubility and implement proactive solubility enhancement techniques. The current research proposes three predictive models based on four solubility datasets (ESOL, AQUA, PHYS, OCHEM), encompassing 3942 unique molecules. Three different molecular representations were obtained, including electrostatic potential (ESP) maps, molecular graph, and tabular features (extracted from ESP maps and tabular Mordred descriptors). We conducted 3942 DFT calculations to acquire ESP maps and extract features from them. Subsequently, we applied two deep learning models, EdgeConv and Graph Convolutional Network (GCN), to the point cloud (ESP) and graph modalities of molecules. In addition, we utilized a random forest-based feature selection on tabular features, followed by mapping with XGBoost. A t-SNE analysis visualized chemical space across datasets and unique molecules, providing valuable insights for model evaluation. The proposed machine learning (ML)-based models, trained on 80% of each dataset and evaluated on the remaining 20%, showcased superior performance, particularly with XGBoost utilizing the extracted and selected tabular features. This yielded average test data Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R-squared (R²) values of 0.458, 0.613, and 0.918, respectively. Furthermore, an ensemble of the three models showed improvement in error metrics across all datasets, consistently outperforming each individual model. This Ensemble model was also tested on the Solubility Challenge 2019, achieving an RMSE of 0.865 and outperforming 37 models with an average RMSE of 1.62. Transferability analysis of our work further indicated robust performance across different datasets. Additionally, SHAP explainability for the feature-based XGBoost model provided transparency in solubility predictions, enhancing the interpretability of the results.

We would like to take this opportunity to thank all of Digital Discovery’s reviewers for helping to preserve quality and integrity in chemical science literature. We would also like to highlight the Outstanding Reviewers for Digital Discovery in 2023.

Reaction screening and high-throughput experimentation (HTE) coupled with liquid chromatography (HPLC and UHPLC) are becoming more important than ever in synthetic chemistry. With a growing number of experiments, it is increasingly difficult to ensure correct peak identification and integration, especially due to unknown side components which often overlap with the peaks of interest. We developed an improved version of the MOCCA Python package with a web-based graphical user interface (GUI) for automated processing of chromatograms, including baseline correction, intelligent peak picking, peak purity checks, deconvolution of overlapping peaks, and compound tracking. The individual automatic processing steps have been improved compared to the previous version of MOCCA to make the software more dependable and versatile. The algorithm accuracy was benchmarked using three datasets and compared to the previous MOCCA implementation and published results. The processing is fully automated with the possibility to include calibration and internal standards. The software supports chromatograms with photo-diode array detector (DAD) data from most commercial HPLC systems, and the Python package and GUI implementation are open-source to allow addition of new features and further development.