Edward O. Pyzer-Knapp

A significant challenge in computational chemistry is developing approximations that accelerate ab initio methods while preserving accuracy. Machine learning interatomic potentials (MLIPs) have emerged as a promising solution for constructing atomistic potentials that can be transferred across different molecular and crystalline systems. Most MLIPs are trained only on energies and forces in vacuum, while an improved description of the potential energy surface could be achieved by including the curvature of the potential energy surface. We present Hessian QM9, the first database of equilibrium configurations and numerical Hessian matrices, consisting of 41,645 molecules from the QM9 dataset at the ωB97x/6-31G* level. Molecular Hessians were calculated in vacuum, as well as water, tetrahydrofuran, and toluene using an implicit solvation model. To demonstrate the utility of this dataset, we show that incorporating second derivatives of the potential energy surface into the loss function of a MLIP significantly improves the prediction of vibrational frequencies in all solvent environments, thus making this dataset extremely useful for studying organic molecules in realistic solvent environments for experimental characterization.

Graph neural networks have recently met huge success in various inference tasks including materials property prediction amongst many others. Nevertheless, having an inherently locally-based representation capacity as they do, global representation of materials' structures can only only be achieved by expanding the model complexity which in turn scales up training times and memory consumption. In this work we focus on efficiently capturing global interactions ``in-model'', through long-range edge attentions with minimal memory footprints. We introduce a novel ``contextual'' message passing scheme that better captures global interactions by attending on edges from both the local and global environment of each node in an edge-update fashion. The performance of the proposed model (LiCOMPGNN) is tested on a diverse set of materials property prediction benchmarks and demonstrates competitiveness against state-of-the-art models in several prediction tasks whilst being an order of magnitude smaller in terms of trainable parameters. We further augment the framework to a multiplex graph setting for solid-state data with reciprocal space features taken into account in a multimodal message passing regime. We demonstrate the representation capacity of the proposed variant along with others in maintaining supercell invariance in crystalline property prediction tasks.

The emergence of Large Language Models (LLMs) have fundamentally altered the way we interact with digital systems and have led to the pursuit of LLM powered AI agents to assist in daily workflows. LLMs, whilst powerful and capable of demonstrating some emergent properties, are not logical reasoners and often struggle to perform well at all sub-tasks carried out by an AI agent to plan and execute a workflow. While existing studies tackle this lack of proficiency by generalised pretraining at a huge scale or by specialised fine-tuning for tool use, we assess if a system comprising of a coalition of pretrained LLMs, each exhibiting specialised performance at individual sub-tasks, can match the performance of single model agents. The coalition of models approach showcases its potential for building robustness and reducing the operational costs of these AI agents by leveraging traits exhibited by specific models. Our findings demonstrate that fine-tuning can be mitigated by considering a coalition of pretrained models and believe that this approach can be applied to other non-agentic systems which utilise LLMs.

Conjugated organic photoredox catalysts (OPCs) can promote a wide range of chemical transformations. It is challenging to predict the catalytic activities of OPCs from first principles, either by expert knowledge or by using a priori calculations, as catalyst activity depends on a complex range of interrelated properties. Organic photocatalysts and other catalyst systems have often been discovered by a mixture of design and trial and error. Here we report a two-step data-driven approach to the targeted synthesis of OPCs and the subsequent reaction optimization for metallophotocatalysis, demonstrated for decarboxylative sp³–sp² cross-coupling of amino acids with aryl halides. Our approach uses a Bayesian optimization strategy coupled with encoding of key physical properties using molecular descriptors to identify promising OPCs from a virtual library of 560 candidate molecules. This led to OPC formulations that are competitive with iridium catalysts by exploring just 2.4% of the available catalyst formulation space (107 of 4,500 possible reaction conditions).

By drawing on similarities between energy landscapes and model response surfaces we gain new insight into model performance, even in the absence of data.

In this work, we outline how methods from the energy landscapes field of theoretical chemistry can be applied to study machine learning models. Various applications are found, ranging from interpretability to improved model performance.

The story of machine learning in general, and its application to molecular design in particular, has been a tale of evolving representations of data. Understanding the implications of the use of a particular representation -- including the existence of so-called `activity cliffs' for cheminformatics models -- is the key to their successful use for molecular discovery. In this work we present a physics-inspired methodology which exploits analogies between model response surfaces and energy landscapes to richly describe the relationship between the representation and the model. From these similarities, a metric emerges which is analogous to the commonly used frustration metric from the chemical physics community. This new property shows state-of-the-art prediction of model error, whilst belonging to a novel class of roughness measure that extends beyond the known data allowing the trivial identification of activity cliffs even in the absence of related training or evaluation data.

The ever-increasing capability of computational methods has resulted in their general acceptance as a key part of the materials design process. Traditionally this has been achieved using a so-called computational funnel, where increasingly accurate - and expensive – methodologies are used to winnow down a large initial library to a size which can be tackled by experiment. In this paper we present an alternative approach, using a multi-output Gaussian process to fuse the information gained from both experimental and computational methods into a single, dynamically evolving design. Common challenges with computational funnels, such as mis-ordering methods, and the inclusion of non-informative steps are avoided by learning the relationships between methods on the fly. We show this approach reduces overall optimisation cost on average by around a factor of three compared to other commonly used approaches, through evaluation on three challenging materials design problems.