Digital discovery最新文献

Organometallic complexes are ubiquitous in numerous technological applications, and in particular in homogeneous catalysis. Optimization of such complexes for specific applications is challenging due to the large variety of possible metal–ligand combinations and ligand–ligand interactions. Here we present OM-Diff, an inverse-design framework based on a diffusion generative model for in silico design of such complexes. Due to the importance of the spatial structure of a catalyst, the model operates on all-atom (including H) representations in 3D space. To handle the symmetries inherent to that data representation, OM-Diff combines an equivariant diffusion model with an equivariant property predictor. The diffusion model generates ligands conditioned on a specified metal-center, while the property predictor guides the generation towards novel complexes with desired properties. We demonstrate the potential of OM-Diff by designing optimized catalysts for a family of cross-coupling reactions, and validating a selection of novel proposed compounds with DFT calculations.

Language models trained on molecular string representations have shown strong performance in predictive and generative tasks. However, practical applications require not only making accurate predictions, but also explainability – the ability to explain the reasons and rationale behind the predictions. In this work, we explore explainability for a chemical language model by adapting a transformer-specific and a model-agnostic input attribution technique. We fine-tune a pretrained model to predict aqueous solubility, compare training and architecture variants, and evaluate visualizations of attributed relevance. The model-agnostic SHAP technique provides sensible attributions, highlighting the positive influence of individual electronegative atoms, but does not explain the model in terms of functional groups or explain how the model represents molecular strings internally to make predictions. In contrast, the adapted transformer-specific explainability technique produces sparse attributions, which cannot be directly attributed to functional groups relevant to solubility. Instead, the attributions are more characteristic of how the model maps molecular strings to its latent space, which seems to represent features relevant to molecular similarity rather than functional groups. These findings provide insight into the representations underpinning chemical language models, which we propose may be leveraged for the design of informative chemical spaces for training more accurate, advanced and explainable models.

Accurate prediction of diverse chemical properties is crucial for advancing molecular design and materials discovery. Here we present a versatile approach that uses the intermediate information of a universal neural network potential as a general-purpose descriptor for chemical property prediction. Our method is based on the insight that by training a sophisticated neural network architecture for universal force fields, it learns transferable representations of atomic environments. We show that transfer learning with graph neural network potentials such as M3GNet and MACE achieves accuracy comparable to state-of-the-art methods for predicting the NMR chemical shifts by using quantum machine learning as well as a standard classical regression model, despite the compactness of its descriptors. In particular, the MACE descriptor demonstrates the highest accuracy to date on the ¹³C NMR chemical shift benchmarks for drug molecules. This work provides an efficient way to accurately predict properties, potentially accelerating the discovery of new molecules and materials.

The rapid growth of automated and autonomous instrumentation brings forth opportunities for the co-orchestration of multimodal tools that are equipped with multiple sequential detection methods or several characterization techniques to explore identical samples. This is exemplified by combinatorial libraries that can be explored in multiple locations via multiple tools simultaneously or downstream characterization in automated synthesis systems. In co-orchestration approaches, information gained in one modality should accelerate the discovery of other modalities. Correspondingly, an orchestrating agent should select the measurement modality based on the anticipated knowledge gain and measurement cost. Herein, we propose and implement a co-orchestration approach for conducting measurements with complex observables, such as spectra or images. The method relies on combining dimensionality reduction by variational autoencoders with representation learning for control over the latent space structure and integration into an iterative workflow via multi-task Gaussian Processes (GPs). This approach further allows for the native incorporation of the system's physics via a probabilistic model as a mean function of the GPs. We illustrate this method for different modes of piezoresponse force microscopy and micro-Raman spectroscopy on a combinatorial Sm-BiFeO₃ library. However, the proposed framework is general and can be extended to multiple measurement modalities and arbitrary dimensionality of the measured signals.

Exploiting crystallographic data repositories for large-scale quantum chemical computations requires the rapid and accurate extraction of the molecular structure, charge and spin from the crystallographic information file. Here, we develop a general approach to assign the ground state spin of transition metal complexes, in complement to our previous efforts on determining metal oxidation states and bond order within the cell2mol software. Starting from a database of 31k transition metal complexes extracted from the Cambridge Structural Database with cell2mol, we construct the TM-GSspin dataset, which contains 2063 mononuclear first row transition metal complexes and their computed ground state spins. TM-GSspin is highly diverse in terms of metals, metal oxidation states, coordination geometries, and coordination sphere compositions. Based on TM-GSspin, we identify correlations between structural and electronic features of the complexes and their ground state spins to develop a rule-based spin state assignment model. Leveraging this knowledge, we construct interpretable descriptors and build a statistical model achieving 98% cross-validated accuracy in predicting the ground state spin across the board. Our approach provides a practical way to determine the ground state spin of transition metal complexes directly from crystal structures without additional computations, thus enabling the automated use of crystallographic data for large-scale computations involving transition metal complexes.

Asymmetric catalysis plays a crucial role in advancing medicine and materials science. However, the prevailing experiment-driven methods for catalyst evaluation are both resource-heavy and time-consuming. To address this challenge, we present CatScore – a learning-centric metric designed for the automatic evaluation of catalyst design models at both instance and system levels. This approach harnesses the power of deep learning to predict product selectivity as a function of reactants and the proposed catalyst. The predicted selectivity serves as a quantitative score, enabling a swift and precise assessment of a catalyst's activity. On an instance level, CatScore's predictions correlate closely with experimental outcomes, demonstrating a Spearman's ρ = 0.84, which surpasses the density functional theory (DFT) based linear free energy relationships (LFERs) metric with ρ = 0.55 and round-trip accuracy metrics at ρ = 0.24. Importantly, when ranking catalyst candidates, CatScore achieves a mean reciprocal ranking significantly superior to traditional LFER methods, marking a considerable reduction in labor and time investments needed to find top-performing catalysts.

Informatics-driven approaches, such as machine learning and sequential experimental design, have shown the potential to drastically impact next-generation materials discovery and design. In this perspective, we present a few guiding principles for applying informatics-based methods towards the design of novel nuclear waste forms. We advocate for adopting a system design approach, and describe the effective usage of data-driven methods in every stage of such a design process. We demonstrate how this approach can optimally leverage physics-based simulations, machine learning surrogates, and experimental synthesis and characterization, within a feedback-driven closed-loop sequential learning framework. We discuss the importance of incorporating domain knowledge into the representation of materials, the construction and curation of datasets, the development of predictive property models, and the design and execution of experiments. We illustrate the application of this approach by successfully designing and validating Na- and Nd-containing phosphate-based ceramic waste forms. Finally, we discuss open challenges in such informatics-driven workflows and present an outlook for their widespread application for the cleanup of nuclear wastes.

The absence of computational methods to predict stressor-specific degradation susceptibilities represents a significant and costly challenge to the introduction of new materials into applications. Here, a machine-learning framework is developed that predicts stressor-specific stability scores from computationally generated reaction data. The thermal degradation of alkanes was studied as an exemplary system to demonstrate the approach. The half-lives of ∼32k alkanes were simulated under pyrolysis conditions using 59 model reactions. Using a hinge-loss function, these half-life data were used to train machine learning models to predict a scalar representing the relative stability based only on the molecular graph. These models were successful in transferability case studies using distinct training and testing splits to recapitulate known stability trends with respect to the degree of branching and alkane size. Even the simplest models showed excellent performance in these case studies, demonstrating the relative ease with which thermal stability can be learned. The stability score is also shown to be useful in a design study, where it is used as part of the objective function of a genetic algorithm to guide the search for more stable species. This work provides a framework for converting kinetic reaction data into stability scores that provide actionable design information and opens avenues for exploring more complex chemistries and stressors.

The vast array of shapes achievable through modern nanofabrication technologies presents a challenge in selecting the most optimal design for achieving a desired optical response. While data-driven techniques, such as deep learning, hold promise for inverse design, their applicability is often limited as they typically explore only smaller subsets of the extensive range of shapes feasible with nanofabrication. Additionally, these models are often regarded as ‘black boxes,’ lacking transparency in revealing the underlying relationship between the shape and optical response. Here, we introduce a methodology tailored to address the challenges posed by large, complex, and diverse sets of nanostructures. Specifically, we demonstrate our approach in the context of periodic silicon metasurfaces operating in the visible wavelength range, considering large and diverse shape set variations. Our paired variational autoencoder method facilitates the creation of rich, continuous, and parameter-aligned latent space representations of the shape–response relationship. We showcase the practical utility of our approach in two key areas: (1) enabling multiple-solution inverse design and (2) conducting sensitivity analyses on a shape's optical response to nanofabrication-induced distortions. This methodology represents a significant advancement in data-driven design techniques, further unlocking the application potential of nanophotonics.

The unique aspect of hybrid perovskites is their tunability, allowing the engineering of the bandgap via substitution. From the application viewpoint, this allows creation of tandem cells between perovskites and silicon, or two or more perovskites, with associated increase of efficiency beyond the single-junction Shockley–Queisser limit. However, the concentration dependence of the optical bandgap in hybrid perovskite solid solutions can be non-linear and even non-monotonic, as determined by band alignments between endmembers, presence of defect states and Urbach tails, and phase separation. Exploring new compositions brings forth the joint problem of the discovery of the composition with the desired band gap and establishing the physical model of the band gap concentration dependence. Here we report the development of the experimental workflow based on structured Gaussian Process (sGP) models and custom sGP (c-sGP) that allow the joint discovery of the experimental behavior and the underpinning physical model. This approach is verified with simulated datasets with known ground truth and was found to accelerate the discovery of experimental behavior and the underlying physical model. The d/c-sGP approach utilizes a few calculated thin film bandgap data points to guide targeted explorations, minimizing the number of thin film preparation steps. Through iterative exploration, we demonstrate that the c-sGP algorithm that combined 5 bandgap models converges rapidly, revealing a relationship in the bandgap diagram of MA_1−xGA_xPb(I_1−xBr_x)₃. This approach offers a promising method for efficiently understanding the physical model of band gap concentration dependence in binary systems, and this method can also be extended to ternary or higher dimensional systems.