Themed collection Data-driven discovery in the chemical sciences

Syntheseus provides reference models and search algorithms as well as metrics to evaluate and improve synthesis planning tools.

We discuss how machine learning researchers view and approach problems in chemistry and provide our considerations for maximizing impact when researching machine learning for chemistry.

A set of 6 filters based on chemical rules, human intuition, and practical constraints are developed to screen for synthesizable compounds. When applied to over 100 000 generated compounds in 60 phase diagrams, 27 are identified as possibly.

The InChI standard facilitates chemical compound identification across platforms, with v1.07 fixing numerous issues and enhancing transparency via GitHub. This update aims to better represent molecular inorganic compounds, addressing previous limitations.

Machine-learned regression or classification models built from historical materials synthesis datasets have limited utility in guiding the predictive synthesis of novel materials, but anomalous recipes can inspire surprising new synthesis strategies.

Improving accessibility of data-driven optimisation for chemical tasks via a graphical user interface.

We have built the first transformers trained on the property-to-molecular-graph task, which we dub “large property models”. A key ingredient is supplementing these models during training with relatively basic but abundant chemical property data.

We present an accurate and data-efficient protocol for fine-tuning the MACE-MP-0 foundational model for a given system. Our model achieves kJ/mol in predicting sublimation enthalpies and below 1% error in the density of ice polymorphs.

Assessment of uncertainty estimates of neural fingerprint-based models by comparing deep learning-based models with combinations of neural fingerprints and classical machine learning algorithms that employ established uncertainty calibration methods.

We demonstrate the wide utility of prediction rigidities, a family of metrics derived from the loss function, in understanding the robustness of machine learning (ML) model predictions.

We investigate three related questions: can we identify the sequence determinants which lead to protein self interactions and phase separation; can we understand and design new sequences which selectively bind to protein condensates?; can we design multiphasic condensates?

The advent of larger datasets in materials science poses unique challenges in modeling, infrastructure, and data diversity and quality.

We evaluate FlanT5 and ByT5 across tokenisation, pretraining, finetuning and inference and benchmark their impact on organic reaction prediction tasks.

We introduce a strategy to train machine learning potentials using MACE, an equivariant message-passing neural network, for metal–ligand complexes in explicit solvents.

We derive maximum and realistic performance bounds based on experimental errors for commonly used machine learning (ML) datasets for regression and classification and compare them to the reported performance of ML models.

We adopt the latest approaches in equivariant graph neural networks to develop a model that can predict the full dielectric tensor of crystals, discovering crystals with almost isotropic connectivity but highly anisotropic dielectric tensors.

Active Thermochemical Tables (ATcT) are employed to resolve existing inconsistencies surrounding the thermochemistry of glycine and produce accurate enthalpies of formation for this system.

Leveraging natural language processing models including transformers, we curate four distinct datasets: tmCAT for catalysis, tmPHOTO for photophysical activity, tmBIO for biological relevance, and tmSCO for magnetism.

We demonstrate the reliability and scalability of computational crystal structure prediction (CSP) methods for small, rigid organic molecules by performing in-depth CSP investigations for over 1000 such compounds.

We integrate generative machine learning with heuristic crystal structure prediction in FUSE. The combined result shows superior performance over both components, accelerating the pace at which we will be able to predict and discover new compounds.

Ephemeral Data-Derived Potential (EDDP)-driven long high-temperature anneals combined with AIRSS, termed as hot-AIRSS, enable the exploration of low-energy configurations of complex materials.

New hypothetical compounds are reported in a collection of online databases. By combining active learning with density-functional theory calculations, this work screens through such databases for materials with optical applications.

Knowledge distillation worked to improve the neural network potential for organic molecular crystals.

We enumerate binary, ternary, and quaternary element and species combinations and present a two-dimensional representation of inorganic crystal chemical space, labelled according to whether the combinations pass standard chemical filters and if they appear in known databases.