Artifical Neural Networks for Processing Velocity Map Images*

We present the first demonstration of artificial neural networks (ANNs) for the removal of Poissonian noise in velocity map imaging (VMI) measurements with very low overall counts. The approach is successfully applied to both simulated and real experimental data relating to the detection of photoions/photoelectrons in unimolecular photochemical dynamics studies. Our results reveal an excellent level of performance, with the ANNs transforming images that are unusable for any form of quantitative analysis into statistically reliable data with an impressive similarity to benchmark references. We also then present a second ANN strategy that we refer to as Arbitrary Image Reinflation (AIR), developed for reconstructing 3D photoproduct distributions from a single 2D projection. The AIR approach is demonstrated for distributions that possess cylindrical symmetry about an axis parallel to the imaging plane and, importantly, those that also do not. Given the widespread use of VMI methods within the chemical dynamics community, we anticipate that the use of ANNs for data processing has significant potential impact – particularly, for example, when working in the limit of very low absorption/photoionization cross-sections, or when attempting to reliably extract subtle image features originating from phenomena such as photofragment vector correlations or photoelectron circular/elliptical dichroism.