Advanced Neural Network Methods for Atomistic Materials Chemistry

机器学习电位（mlp）正在成为越来越有吸引力的工具，用于模拟分子和材料；在简单的域内，它们的精度已经收敛到DFT方法，而计算成本则低了几个数量级。这使得对更大的系统和更长的时间尺度进行彻底的研究成为可能，所有这些都接近化学精度。正是在这些时间和长度尺度上，许多与材料性质有关的微观过程发生了。个人mlp的普遍性也在提高。然而，对于复杂的、反应性的和多组分的材料，它们的开发仍然很困难。为了模拟，更重要的是，理解催化过程，因此需要新的方法。该项目将寻求将计算机视觉和自然语言处理技术应用于化学领域。具体而言，将探讨迁移学习（TL）方法的最新进展；迄今为止，化学文献中只使用了有限的TL技术，主要集中在分子数据集上。本研究将寻求应用更多最新的语言训练技术，以提高mlp的准确性、训练效率和通用性。该项目旨在通过创建强大的mlp来提高对催化过程的理解，预计将为更广泛的计算化学和材料科学研究界提供方法和研究数据。为复杂和反应性系统创建mlp具有挑战性。该项目将制定培训mlp的新方法。这些方法将被应用于研究原子尺度的催化过程，有助于进一步了解这些系统的动力学。该项目属于EPSRC“计算与理论化学”和“催化”研究领域，在“物理科学”主题下。该项目的目标与该主题的既定战略相一致，即应对“依赖基础科学解决的许多社会和经济挑战”。关于第一个研究领域，该项目具有明确的基础性质：它将导致新的计算方法的发展，用于在原子尺度上描述系统（期望在此过程中解决有关原子性质“学习”的基本问题），然后该方法将用于研究这些系统随时间的动力学。此外，通过将计算机科学领域的发展与化学和材料科学的新兴应用相结合，该项目符合该领域实现“跨学科研究”的关键目标。至于第二个研究领域，从长远来看，该项目有望为“结构和动力学研究以理解催化机制”做出贡献。

Machine Learned Potentials (MLPs) are becoming increasingly attractive tools with which to simulate both molecules and materials; in simple domains, their accuracy has converged to that of DFT methods, while their computational cost is orders of magnitudes lower. This has allowed thorough investigation into larger systems and for longer timescales, all at near chemical accuracy. It is on these time- and length-scales that many microscopic processes which are relevant for materials properties take place. The generality of individual MLPs is also improving. However, they are still difficult to develop for complex, reactive and multicomponent materials. In order to simulate and, more importantly, understand catalytic processes, new methodologies are therefore required. This project will seek to adapt techniques from computer vision and natural language processing for use in the chemical domain. Concretely, recent advances in transfer learning (TL) methodologies will be explored; only a limited set of TL techniques have been used in the chemical literature to date, predominantly with a focus on molecular datasets. This research will seek to apply more recent TL techniques with the goal of improving the accuracy, efficiency of training and generalisability of MLPs. With an aim to improve understanding of catalytic processes by creating powerful MLPs, the project is expected to lead to methods and research data which will be made available to the wider computational chemistry and materials science research community. Creating MLPs for complex and reactive systems is challenging. This project will develop a new approach to training MLPs. These approaches will then be applied to study atomic-scalecatalytic processes, helping to further understand the dynamics of these systems. This project falls within the EPSRC "Computational and theoretical chemistry" and "Catalysis" research areas, under the "Physical sciences" theme. The project's aims align with this theme's stated strategy of meeting "the many societal and economic challenges that rely on fundamental science for solutions". With regards to the first research area, this project is of clear fundamental nature: it will lead to the development of new computational methodology with which to describe systems at an atomic scale (with an expectation to address fundamental questions about the "learning" of atomistic properties in the process), and this methodology will then be used to study the dynamics of these systems over time. Additionally, by aiming to combine developments in the field of computer science with emerging applications in chemistry and material science, this project aligns with the area's key goal of reaching "across disciplines with research". With regards to the second research area, this project is expected to contribute, in the long term, to "structural and kinetic studies to understand catalytic mechanisms."