Application of machine learning to condensed matter physics

点击翻译按钮获取中文摘要

Machine learning has already proven to be one of the most influential discoveries of our time. It has found widespread applications in almost every scientific field, becoming an essential predictor of important properties of condensed matter materials. Developing purpose-built machine learning algorithms for studying these properties would streamline research in the search for new desirable materials from the extremely vast number of possible constructions. The proposed research will pioneer an interdisciplinary approach, utilising both novel machine learning techniques and quantum computing algorithms to study condensed matter systems. This combination will be well-suited to yield insights into the quantum mechanical behavior of these materials.Condensed matter physics is the study and modeling of solid-state materials using quantum mechanical descriptions. Due to many quantum interactions, finding exact solutions to these descriptions is not feasible. There have been attempts to develop computational methods that can closely approximate these emergent properties so that new materials can be designed with desirable features. Neural networks appear to be an ideal method to analyze these systems. They are highly effective at efficiently searching through large search spaces, making them applicable to condensed matter problems, which feature a broad array of possible material constructions.Purpose-built condensed matter neural networks could provide significant insights into the physics of condensed matter systems. The proposed research aims to redesign neural network algorithms with the guiding principle of removing implicit assumptions already identified in the current formulations. This will increase the symmetry properties of networks, consequently making it particularly applicable to condensed matter systems that feature many analogous symmetries. These redesigns appear to be widely applicable, even outside the domain of condensed matter applications; the resulting research should be influential, with many functions that can be overhauled using this symmetry principle.This project will consist of a novel and large overhaul of many aspects of current neural networks, with far-reaching improvements across many of their applications. Particularly, the approach of removing implicit assumptions and consequently introducing symmetries will have a substantial impact in condensed matter physics and quantum machine learning, with improvements to quantum machine learning aiding further advances in condensed matter modeling.Below is a summary of the expected outcomes of this work:New highly efficient neural network functions exhibiting symmetries applicable to condensed matter systems. These should allow the network to represent and manipulate data more effectively.Improved predictions of the properties of condensed matter systems.Potential to design materials by specifying the desired characteristics.Insight into the nature of how neural networks embed and manipulate data.Interfaces with quantum machine learning and new activation functions that better convey quantum mechanical probabilities.Further advances in recurrent neural network models, with improved memory efficiency. This should allow for an iterative approach to modeling condensed matter systems.