Machine learning techniques in Materials Science

Atomistic Machine Learning between Physics and Data

Professor Michele Ceriotti, EPFL

Machine Learning-based Interatomic Potentials for Complex Materials - Taking Stock and Thinking Ahead

Dr. Gabriele Sosso, University of Warwick

Thursday 25th April 2019
Time: 5pm
Venue: Lecture Room G20, Royal School of Mines, Imperial College London
Contact: Ms Hafiza Bibi
Tel: 020 7594 7252

Atomistic Machine Learning between Physics and Data

Prof. Michele Ceriotti (EPFL)

Abstract: Statistical regression techniques have become very fashionable as a tool to predict the properties of systems at the atomic scale, sidestepping much of the computational cost of first-principles simulations and making it possible to perform simulations that require thorough statistical sampling without compromising on the accuracy of the electronic structure model. 

In this talk I will argue how data-driven modelling can be rooted in a mathematically rigorous and physically-motivated framework, and how this is beneficial to the accuracy and the transferability of the model. I will also highlight how machine learning - despite amounting essentially at data interpolation - can provide important physical insights on the behavior of complex systems, on the synthesizability and on the structure-property relations of materials.

I will give examples concerning all sorts of atomistic systems, from semiconductors to molecular crystals, and properties as diverse as drug-protein interactions, dielectric response of aqueous systems and NMR chemical shielding in the solid state.


Machine Learning-based Interatomic Potentials for Complex Materials - Taking Stock and Thinking Ahead

Dr. Gabriele Sosso (University of Warwick)

Abstract: The next generation of functional materials for applications as diverse as quantum computing or aerospace engineering is likely to be characterised by an ever-increasing level of complexity: structural complexity (nano-pattering, porous materials…) and chemical complexity (multicomponent alloys, intercalation compounds). Atomistic simulations are bound to play a key role in unravelling the microscopic details of these classes of materials, but the sheer level of complexity involved requires in many cases the accuracy of first principles (e.g. density functional theory based) calculations as well as the large/long length/time scales available by classical (e.g. empirical force fields-based) simulations. Machine Learning-based Interatomic Potentials (MLIPs) represent a sensible way forward to bring together the best of both (first principles and classical) worlds, and indeed in the last few years we have witnessed a number of very successful applications. In this talk, I will discuss the curious case of phase change materials (promising candidates for the next generation of non-volatile memories), focussing on a number of their different functional properties we have characterised via the usage of a MLIP within the last few years. In addition, I will offer a rather subjective perspective on the many open questions currently hampering the field: the need to rapidly construct reliable reference datasets (and being able to expand them quickly to incorporate additional structural and chemical complexity), the choice of the descriptors we use to represent said datasets into a machine learning-digestible form, the option of harness MLIPs-inspired frameworks to further drug design and discovery, and, perhaps controversially, whether or not to opt for the construction of a MLIP in the first place. 



Follow @tyc_london for updates from the Thomas Young Centre.