BasisFlow: Machine learning for tailor made quantum chemistry

点击翻译按钮获取中文摘要

Computer modelling of chemical reactions is used to guide and interpret experiments, allowing scientists to explore exciting new areas of chemistry as they are discovered. But, the time taken for the calculations limits how big the chemical system can be and how accurate the results are. We can currently carry out calculations on roughly 50-100 atom systems and are reaching the limits of what we can do, but we need to able to consider hundreds or thousands of atoms to make developments that affect wider society: for example, transforming industrial chemistry to be more environmentally friendly, aiding discovery of new antibiotics, or helping to design bio-degradable plastics.In this research project we are going to address part of the puzzle of how to carry out accurate calculations on much larger chemical systems. We will do this by developing new theoretical tools that are more accurate and produce results more quickly than the existing alternatives. We will also develop a computer program that will allow other scientists working with computational chemistry to tailor their theoretical tools to the specific applications they are working on, which will make it easier for them to keep pace with the latest discoveries in experimental chemistry and its interfaces with physics, materials science and biology. To do this we will use techniques from the field of artificial intelligence to create a step change in how we approach the challenge of developing new theoretical chemistry tools and techniques. By applying similar methods as those used by large technology companies in, for example, self-driving cars, we will be able to consider large data sets of molecules and their properties, which will mean our tools will be optimised for hundreds of thousands more molecules than would be possible with existing methods.Contemporary artificial intelligence relies on having large amounts of high-quality data, which is why part of this research project will focus on learning the best ways to generate the data about molecules, and the chemistry they undergo, for the purposes of advancing computational chemistry. Not only will we construct and use new data sets in our own development of computational tools, we will also establish a workflow that allows other scientists to produce compatible data sets and move this research into many other areas of chemistry. The tools and software developed in this work will be made available to all, but will be of particular interest to chemists, both those working at universities and in industry, who use computational chemistry to accompany what happens in the lab. Accelerating their analysis of results will allow them to make better predictions and interpretations, ultimately guiding experiment in the right direction. Increasing the size of chemical system that is feasible for computers to model can also redefine how we use theoretical methods to learn more about chemistry, and the role it plays in our modern world.