Deep Neural Network Predicts Time-Resolved Spectroscopic Data
A team of researchers from the University of York and Newcastle University have trialled an ‘on-the-fly’ deep neural network (DNN) training protocol that can be used to simulate the spectroscopic signals from ultrafast molecular processes.
The latest high-brilliance light sources at particle accelerator facilities have transformed the kind of science that it is possible to do today, unlocking a universe of ultrafast high-resolution time-resolved experimental techniques that are able reveal fine details about the excited-state dynamics of molecules and materials on the atomic scales of length and time.
Many of these techniques (eg. ultrafast (multidimensional) spectroscopy, stimulated Raman, and electron and X-ray scattering experiments) offer complimentary insights into the coupled nuclear and electronic dynamics that underpin the functional properties of the molecules and materials that are able to be studied. Connecting the experimental observables to the operational chemistry and physics through computational simulations is the best way to understand these physical phenomena and, ultimately, an important step towards harnessing them for practical applications. The complexity of these techniques is such that this is an incredibly tall order, however, and there is real demand for approaches to computational simulation that are fast, accurate, affordable, and generally-applicable enough to act as a ‘limited-experience-necessary’ starting point.
In the new research, Dr Conor Rankine, Prof. Tom Penfold, and Clelia Middleton (the PhD student who led the work) have shown that their XANESNET DNN for X-ray absorption and emission spectroscopy can be used with an ‘on-the-fly’ training protocol, allowing it to analyse a limited quantity of computational simulation data and converge quickly to meaningful solutions. Their machine-learning approach is able to sidestep what would otherwise involve days/weeks of time-consuming quantum-chemical calculations and expert oversight, allowing the experimental observables of an ultrafast excited-state dynamic process to be predicted using only a limited sample of geometries explored in the process.
Speaking about the research, Dr Conor Rankine said: “I’m sure that people will remember (and perhaps be sick of!) the ultrafast ring-opening dynamics of 1,2-dithiane [a small, cyclic disulfide] that we used as a test bed in this research – I presented countless posters and talks on this system when I was a postgraduate student at York! It’s great to see Clelia carrying the torch for these disulfide systems – she’s put a Herculean amount of effort into driving this research forwards and putting this out as her first [PhD] publication, and it couldn’t be in safer hands as far as I’m concerned. Congratulations Clelia!”
This research has been published in Physical Chemistry Chemical Physics.