This study presents a novel approach to understanding the structure–property relationships in non-stoichiometric materials and interstitial alloys using graph neural networks (GNNs). Specifically, we apply the crystal graph convolutional network (CGCNet) to predict the properties of transition-metal carbides, Mo₂C and Ti₂C, and introduce the crystal graph explainer (CGExplainer) enabling model interpretability. CGCNet outperforms traditional human-derived interatomic potential models (IAPs) in prediction accuracy and data efficiency, with significant improvements in the ability to extrapolate properties to larger supercells. Additionally, the CGExplainer tool enables detailed analysis of the relative spatial positioning of atomic ensembles, revealing key atomic arrangements that govern material properties. This work highlights the potential of GNN-based approaches for rapidly discovering complex structure–property relationships and accelerating the design of materials with customized properties, particularly for alloys with variable atomic compositions. Our methodology offers a robust framework for future materials discovery, extending the applicability of GNNs to a broader range of materials systems.

The range of chemical databases available has dramatically increased in recent years, but the reliability and quality of their data are often negatively affected by human-error fidelity. The size of chemical databases can make manual data curation/checking of such sets time consuming; thus, automated tools to help this process are highly desirable. Herein, we propose the use of Graph Neural Networks (GNNs) to identifying potential stereochemical misassignments in the primary asymmetric catalysis literature. Our method relies on the use of an ensemble of GNN models to predict the expected stereoselectivity of exemplars for a particular asymmetric reaction. When the majority of these models do not correlate to the reported outcome, the point is labeled as a possible stereochemical misassignment. Such identified cases are few in number and more easily investigated for their cause. We demonstrate the use of this approach to spot potential literature stereochemical misassignments in the ketone products resulting from catalytic asymmetric 1,4-addition of organoboron nucleophiles to Michael acceptors in two different databases, each one using a different family of chiral ligands (bisphosphine and diene ligands). Our results demonstrate that this methodology is useful for curation of medium-sized databases, speeding this process significantly compared to complete manual curation/checking. In the datasets investigated, human expert checking was reduced to 2.2% and 3.5% of the total data exemplars.

Helix is an open-source, extensible, Python-based software framework to facilitate reproducible and interpretable machine learning workflows for tabular data. It addresses the growing need for transparent experimental data analytics provenance, ensuring that the entire analytical process—including decisions around data transformation and methodological choices—is documented, accessible, reproducible, and comprehensible to relevant stakeholders. The platform comprises modules for standardized data preprocessing, visualization, machine learning model training, evaluation, interpretation, results inspection, and model prediction for unseen data. To further empower researchers without formal training in data science to derive meaningful and actionable insights, Helix features a user-friendly interface that enables the design of computational experiments and inspection of outcomes, including a novel interpretation approach to machine learning decisions using linguistic terms all within an integrated environment.

Homogeneous catalysts enable faster conversions of molecules with higher selectivities (stereo- and regioselectivity) in chemical reactions. Traditionally, catalyst improvements are made through empirical trials, where the catalyst is functionalised by adding, removing or modifying groups within its structure and, subsequently, reevaluating the new catalytic activity. This procedure is not efficient and leads to unsuccessful trials that waste resources. Machine learning (ML) approaches have been proposed to accelerate homogeneous asymmetric catalyst optimization. However, these often lack a general descriptor generation procedure to allow encoding of molecules from a broad region of chemical space. To overcome this, we propose a homogeneous catalyst graph neural network (HCat-GNet) for the prediction of selectivity of catalysts given the SMILES of participant molecules. We demonstrate its use in rhodium-catalyzed asymmetric 1,4-addition (RhCAA), a reaction of major importance in organic synthesis. We benchmark HCat-GNet against traditional ML methods for its ability to predict RhCAA stereoselectivity from two chiral diene ligand two datasets; one for learning and one for final testing. For the learning dataset, both traditional ML and HCat-GNet methods give comparable results. However, when presented with the new unseen test dataset, traditional ML models perform poorly, while HCat-GNet retains a general ability to accurately predict product absolute stereochemistry and reaction stereoselectivity. Furthermore, HCat-GNet allows model interpretability, permitting analysis of the effect of ligand substituents in determining reaction selectivity. HCat-GNet shows greater potential for catalyst optimization than traditional ML, as it allows the use of a non-fixed number of participant molecules to train the model, only requiring the SMILES of the molecules to create graph representations. HCat-GNet allows more general models that accurately extrapolate into unseen regions of chemical space.

The properties of periodic solid-state materials are frequently modified by the inclusion of interstitial atoms deposited pseudo-randomly throughout the crystal lattice. Accurately calculating overall lattice properties currently requires time-intensive, high-level quantum (typically DFT) calculations, often on 1000s of stochastic lattices. 'Interatomic potential models' (IPM) can mitigate such computational burdens by awarding discrete coefficients to common interstitial geometric ensembles within the wider lattice, and then summing these in a simple regression. Human-designed IPMs however typically take years to be developed. Herein, using molybdenum carbide (Mo lattice, C interstitial added atom) as a model, we show a Crystal Graph Neural Network (CGNet) workflow that can derive near-perfect IPMs from <300 DFT inputs in seconds. Our pipeline (CGNet+CGExplainer) demonstrates high potential for generality in exploring lattice structure-property relations rapidly while maintaining accuracy and interpretability.

Polymers are used in a wide range of industries for diverse applications but formulating them to attain desired properties remains a challenging task. Polymer informatics has emerged as a powerful approach for the rational design of materials. Here, we present PolyNet, an open-source Python package that provides an end-to-end workflow for polymer informatics. PolyNet: 1) reads and validates datasets of polymer properties, including the monomer structures and their molar ratios; 2) creates vector- and graph-based representations of the polymers; 3) trains and optimises traditional machine learning (ML) models and graph neural networks (GNNs); 4) compares representations and models statistically; 5) deploys the models to screen in silico libraries of polymers, flagging polymers outside the applicability domain of the models and estimating their expected prediction error; and 6) interprets the models to understand structure–property relationships. PolyNet provides unique features such as: modelling of homopolymers and copolymers with any number of components, using the molar ratio of monomers in both vector- and graph-based representations; acceptance of textual representations of monomers or repeating units; calculation of polyBERT fingerprints and PolyMetriX and RDKit descriptors and fingerprints and accepting standalone user precalculated representations; benchmarking of all representation–model combinations on identical data splits using paired statistical tests; and chemically interpretable explanations through fragment-level attribution for GNNs and SHAP analysis for traditional ML models. Every experiment automatically records its configuration, data splits, hyperparameter search spaces and trained models, supporting full reproducibility in line with FAIR principles. PolyNet can be used through Python scripts, a command-line interface, or a graphical user interface for researchers without programming experience. We exemplify PolyNet by two case studies: glass transition temperature prediction for homopolymers, where cluster-based data splits showed that random splits overestimate model performance, the applicability domain identified less reliable predictions, and the model explanations recovered the known effect of chain mobility on T _g ; and ¹⁹F MRI signal-to-noise ratio prediction for copolymers, where encoding monomer molar ratios is key to accurate predictions.