Accelerating Theoretical Spectroscopy through Machine-learning

Lead Research Organisation: University of Warwick
Department Name: Physics

Abstract

Spectroscopy techniques are some of the most powerful experimental tools we have for learning about crucial processes involving energy and charge transfer at the nanoscale. Such processes are central to the function of practically all electronic devices, eg semiconductors, batteries, and photovoltaics, and of organic and biological systems, such as light-harvesting complexes enabling photosynthesis. Spectroscopic tools are becoming ever more sophisticated and their energy, spatial and temporal resolution are continually improving. For example, Warwick hosts the Warwick Centre for Ultrafast Spectroscopy, a state-of-the-art femtosecond laser facility with experiments ranging from the uv to the THz regime.

However, as tools become more sophisticated it becomes ever-harder to interpret their results. This is where quantum mechanical computational modelling techniques can be enormously beneficial. Theoretical spectroscopy predicts and explains the absorption and emission spectra of materials and molecules. From these, we can for example routinely predict the colour, of a dye purely from its molecular structure and environment. Theoretical spectroscopy can also help understand photostability and reactions initiated by light absorption such as those that cause dyes to degrade in UV light, or what causes DNA damage leading to skin cancer. While state-of-the-art quantum mechanical simulation tools can deliver the properties of electronic excited states, they are very computationally expensive, and ab initio molecular dynamics on excited-state energy surfaces is not feasible for describing processes over timescales of anything more than a few picoseconds. Furthermore, large, complex systems such as heterogeneous interfaces, solvated molecules and biomolecules are a great challenge to traditional methods, whose scaling with system size can be poor.

To address these challenges, research by a UK-based team over the last decade has produced a robust implementation of Linear-Scaling Density Functional Theory in the form of the ONETEP code (www.onetep.org), which can perform quantum mechanical calculations at an unprecedented scale. Recently, Hine and co-workers at Warwick have added functionality for theoretical spectroscopy, including optical absorption, Electron Energy Loss Spectroscopy, and vibration spectroscopy.

In this PhD project we will take these capabilities to a new level by coupling them with advanced machine-learning techniques, potentially accelerating both electronic structure calculations themselves, and dynamics based on potential energy surfaces derived from them. QM simulations can provide training data to a machine learning packages enabling them to learn the excited state potential energy surfaces of a system. A neural-network representation takes the vector of input coordinates, forms molecular descriptors, and outputs the energies and forces required for dynamics, much more rapidly than any QM-based method possibly could. This enables rapid calculation of molecular dynamics trajectories, helping to understand the "fate" of photoexcited molecules in complex environments. Further possible application areas for ML techniques involve directly deriving an "optimal" set of local orbitals for a system from the positions of nearby atoms, accelerating otherwise costly computational optimisation.

The goals of this project are closely aligned with a range of sub-themes in the EPSRC Physical Sciences portfolio (Chemical reaction dynamic, chemical structure, light-matter interactrions, electronic structure and theoretical chemistry) and with the Artificial Intelligence and Research Infrastructure (Software Engineering) themes.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/S022848/1 01/04/2019 30/09/2027
2228390 Studentship EP/S022848/1 01/10/2019 30/09/2023 Carlo Maino