AI for Science –The Case for Materials
Lead Research Organisation:
University College London
Abstract
Many societal and technological challenges can be addressed by discovering and designing novel materials. Materials are employed everywhere, from prosthetics and drug delivery in health care to electronic devices to nuclear waste storage, carbon capture, and catalysis to reduce our impact on the environment. The development and design of new advanced materials can help to tackle challenges in key areas such as catalysis of complex chemical reactions, how charge and heat travel through devices, energy generation and storage (batteries or fuel cells) and recovering waste heat (thermoelectric materials). Computational materials modelling methods are now accurate enough to lead this work, but these simulations use a large fraction of high-performance computers worldwide.
Machine Learning (ML) methods provide efficient ways to model complex phenomena and can be used to accelerate materials simulations, post-processing, and analysis of the results. This enables scalable, cost-effective and responsive computer modelling across materials science, which supports real-time decision making, modelling of larger and more realistic materials (for example, to include complex defects present in real materials that either provides the property of interest or is the source of its degradation), and a greater understanding of the “uncertainties” in the simulation for example by understanding the effects of the computer approximations.
The next generation of supercomputers exploit GPUs, as CPU-only machines are less energy efficient. Many ML methods also run efficiently on GPUs, ultimately making materials simulations using them even more energy- and cost-efficient. We have already shown that this kind of ML-based approach is effective at analysing Small Angle X-ray Scattering (SAXS) data (https://doi.org/10.1021/jacsau.4c00368), complementing computationally intensive Monte Carlo simulations, for predicting materials properties that are usually computed using expensive electronic structure methods (https://doi.org/10.1038/s41524-024-01486-1) and even for generating candidate structures for global optimisation approaches (https://doi.org/10.1038/s41467-024-54639-7).
In this work, we will embed ML methods within the popular CASTEP materials modelling software, which is developed in the UK and has a worldwide userbase, including many industrial R & D teams. We will investigate a variety of different ML methods from deep neural networks to generative AI to determine the most suitable approaches to address both fundamental science questions and specific high-priority applications, as identified by the UK materials research community (e.g. the UKRI high-priority use-cases) and piloted by us in a recent joint CCP9 and CCP-NC feasibility study of predicting the electron density in a material using Machine Learning. We will also investigate where else AI can be exploited to help the materials and molecular modelling community and either immediately act on these, plan future research packages, and/or disseminate our findings to our community. For example, we will (a) link at least one our materials software codes to the AI code already developed by the Alan Turing Institute to enable an easier route for exploiting AI routines; (b) survey where AI is already being exploited within and outside our community; and (c) develop how to capture atomic structural coordinates of predicted clusters reported in early publications as images of ball and stick models and make the structural data available to the community via our web-accessible database for cluster structures.
The main aim is to embed AI and ML methods in materials research tools in a fair and inclusive way, to create user-friendly, high-impact software which brings the advantages of AI to materials research in a responsible and reliable way.
Machine Learning (ML) methods provide efficient ways to model complex phenomena and can be used to accelerate materials simulations, post-processing, and analysis of the results. This enables scalable, cost-effective and responsive computer modelling across materials science, which supports real-time decision making, modelling of larger and more realistic materials (for example, to include complex defects present in real materials that either provides the property of interest or is the source of its degradation), and a greater understanding of the “uncertainties” in the simulation for example by understanding the effects of the computer approximations.
The next generation of supercomputers exploit GPUs, as CPU-only machines are less energy efficient. Many ML methods also run efficiently on GPUs, ultimately making materials simulations using them even more energy- and cost-efficient. We have already shown that this kind of ML-based approach is effective at analysing Small Angle X-ray Scattering (SAXS) data (https://doi.org/10.1021/jacsau.4c00368), complementing computationally intensive Monte Carlo simulations, for predicting materials properties that are usually computed using expensive electronic structure methods (https://doi.org/10.1038/s41524-024-01486-1) and even for generating candidate structures for global optimisation approaches (https://doi.org/10.1038/s41467-024-54639-7).
In this work, we will embed ML methods within the popular CASTEP materials modelling software, which is developed in the UK and has a worldwide userbase, including many industrial R & D teams. We will investigate a variety of different ML methods from deep neural networks to generative AI to determine the most suitable approaches to address both fundamental science questions and specific high-priority applications, as identified by the UK materials research community (e.g. the UKRI high-priority use-cases) and piloted by us in a recent joint CCP9 and CCP-NC feasibility study of predicting the electron density in a material using Machine Learning. We will also investigate where else AI can be exploited to help the materials and molecular modelling community and either immediately act on these, plan future research packages, and/or disseminate our findings to our community. For example, we will (a) link at least one our materials software codes to the AI code already developed by the Alan Turing Institute to enable an easier route for exploiting AI routines; (b) survey where AI is already being exploited within and outside our community; and (c) develop how to capture atomic structural coordinates of predicted clusters reported in early publications as images of ball and stick models and make the structural data available to the community via our web-accessible database for cluster structures.
The main aim is to embed AI and ML methods in materials research tools in a fair and inclusive way, to create user-friendly, high-impact software which brings the advantages of AI to materials research in a responsible and reliable way.