Molecular simulation and atomistic modelling have played a key role in developing a deeper understanding and suggesting new developments in materials and molecular science for more than half-a-century. Methods such as density functional theory (DFT) and interatomic potentials have allowed us to quantitatively model the behaviour of systems at different length and time scales. Nonetheless, there has generally been a trade-off between the size/time scales that can be explored and the quantitative accuracy of the methods employed. The recent resurgence in machine learning (ML) and data-driven approaches has promised to close the gap between simulating large systems for long timescales and obtaining accurate predictions of a range properties of interest. At the same time increasing accuracy of predictions of important materials and molecular properties means that ML is now being applied to design chemicals as diverse as small molecule pharamaceuticals, porous frameworks and solid-state battery systems.

The importance of ML for materials and molecular modelling is clearly evidenced by the explosion of research in the area. There is an increasing desire to have researchers who are trained both as molecular simulators and as data scientists; however very few existing curricula prepare students with this combination. This school addresses this need, by offering a solid foundation in the basics of ML, introducing some of the latest relevant developments from computer science, and providing concrete examples of the application of these methods to materials and molecular modelling challenges.

This course is aimed at early career researchers, with a background in materials or molecular science, experience in Python, but little background in ML.

• Introduction to the fundamentals of ML ensuring that all participnats have a solid understanding of the basic principles of building models on data.
• Neural networks, as the basic feature behind many of the most powerful techniques in modern ML. Covering architectures and training via backprop.
• Graph neural networks (GNNs). GNNs, with their inductive bias for graph-based strcuture have been transformative in both molecular and materials machine learning. We will cover the basic graph architecture and provide some recent examples.
• Machine learned interatomic potentials. Covering the lateste developments in MLIPs including generating and tuning models for specific applications. We will look at how these models are usef for both materials and for molecules.
• Generative models. We will cover the concept of a generative model, as well as specific examples - variational autoencoders, generative adversarial networks, stable diffusion.
• On the final day we will specialise into sub-groups to look at the latest developments in ML for materials and moelcules
  • Materials - foundation model interatomic potentials
  • Molecules - De novo design by reinforcement learning

The syllabus will give students a solid grounding in the important statistical concepts and algorithms that underpin machine learning, it will then build on these concepts to demonstate the tools that are at the forefront of applied ML in materials and moleculra modelling.

registration is closed, more details on camml.ac.uk