Biography:

Dr. James A Grant-Jacob is a Senior Research Fellow at the Optoelectronics Research Centre, University of Southampton, who has a diverse research portfolio, including high harmonic generation, laser fabrication, DNA sequencing, imaging, and AI. He has over 170 publications, including collaborations with NASA on laser manufacturing for greenhouse gas detection. Utilizing NVIDIA grants, he’s improved laser-based processes through deep learning. In 2019, he presented his AI-based particle pollution detection research at STEM for BRITAIN in the UK Houses of Parliament. He’s collaborated with companies like Dyson and institutions like Southampton General Hospital.

Abstract:
The structure of pollen has evolved depending on its local environment, competition, and ecology. Aspollen grains are generally of size 10–100 microns with nanometre-scale substructure, scanningelectron microscopy is an important microscopy technique for imaging and analysis. Here, we use style transfer deep learning to allow exploration of latent w-space of scanning electron microscopeimages of pollen grains and show the potential for using this technique to understand evolutionarypathways and characteristic structural traits of pollen grains.

Land plants have their origin in the mid-Palaeozoic era [1]. It is estimated that around 350,000 species offlowering plants have evolved [2], producing pollen grains with sizes ranging from approximately 10–100microns depending on the species, and with substructure on the scale of nanometers [3]. Pollen takes a variety ofshapes, such as trilobal, spherical or hexagonal, and the surface of pollen can have features such as apertures [4]that play a role in mechanisms such as germination and harmomegathy [5]. Imaging of pollen grains is thereforean important technique, as it offers a vital indicator in the health of agriculture crop [6]and the localenvironmental [7].Current understanding of pollen traits and evolution generally involves manual observation, withvarious methods of imaging pollen grains used to identify their external andinternal structure. These includefluorescence microscopy [8]and electron microscopy [9], while modelling using analytical methods has beenused for exploring the formation of pollen grain apertures [10]. However, these approaches can be timeconsuming and modelling biological systems can be challenging. In recent years, developments in graphicsprocessing units (GPUs)and deep learning algorithms have unlocked a new paradigm of large-scale datadriven research [11]. Relevant to palynology, deep learning has been used for pollen identification via visiblelight microscopy [12]and fibre optic-based sensing [13], for imaging of pollen grains from scatteringpatterns [14], and for transforming images of dehydrated pollen grains into images of hydrated pollengrains [15].This manuscript describes the application of deep learning to artificially generate scanning electronmicroscope (SEM)images of pollen grains and then interpolate between the generated images in latent w-spaceto potentially allow the exploration of transformations from one pollen species to another, as conceptualised infigure 1. The latent w-space is a multi-dimensional space that enables close positional representation of data thatare similar in the space externally to the neural network (i.e. training data). Since evolution generally results inincremental changes to the visual appearance of pollen grains, and that the neural network clusters pollen graintypes based on their visual appearance in latent w-space, interpolation and extrapolation across latent w-spacecould be comparable to the evolutionary path between pollen species in a phylogenetic tree. We also show thatw-space latent vectors can be determined that can allow characteristics such as the pollen grain size, to beincreased or decreased in the generated images, potentially allowing the understanding of species morphology evolution.