Please subscribe to our mailing list
to receive the latest information about seminars, conferences, grants, and resources!
Tuesday, June 15 at 2 p.m.
The Physical Sciences Machine Learning Nexus graduate student fellows will make short presentations on their projects and engage with you on question and answer discussions.
Caroline K. Brennan, Chemistry – Data driven analysis of luciferase-luciferin pairs.
Abstract: Bioluminescence imaging (BLI) is a powerful tool that relies on light production from a chemical reaction: the luciferase-catalyzed oxidation of a small molecule substrate (luciferin). Luciferases and luciferins have been used for decades to image biological processes. However, the technology has been limited to imaging one or two cellular processes at a time. Our strategy for developing more multicomponent imaging tools involves generating selective (i.e., "orthogonal") luciferases that process structurally distinct substrates. Toward this end, we have generated 223 mutant luciferases and 23 luciferin analogs, which can potentially provide access to >5 million combinations of orthogonal pairs. We are now identifying the origins of selectivity to rationally design orthogonal pairs. To this end, we acquired photon output data of all luciferases and luciferins. Hierarchical clustering was used to identify compound classes from our screening data. We are now identifying commonly occurring mutations within each compound class, which will be used as a starting point to design new mutant libraries. Ultimately, our findings will further the understanding of our collection of luciferase-luciferin pairs.
Alexander M. Broughton, Physics & Astrononomy – Modeling Gamma-Ray Emission with Deep learning and Implications for Dark Matter.
Abstract: There have been a number of theoretical predictions that self-interacting dark matter would emit a smooth distribution of gamma-rays close to the galactic center (GC). Previous analyses discovered a statistically significant excess of unexplained gamma radiation near the GC after removing background sources, such as galaxies, AGN, and cosmic ray interactions with hydrogen gas. However, our research shows that poor modeling of the hydrogen gas component due to radiative reabsorption can cause a non-Poissonian template fit (NPTF) to misidentify point-source components in a potentially smooth dark matter background. Our research team was able to build a convolution neural network that more accurately predicts the spatial morphology of hydrogen gas by training on datasets of less dense tracer gases, like CO-12 and its isotopologues CO-16/18. This algorithm can be used to improve modeling of gamma-ray emission in extended regions of the sky around the GC.
Kevin Bui, Mathematics – Weighted Anisotropic-Isotropic Total Variation for Effective Image Segmentation.
Abstract: Based on the nonconvex regularization penalty, the weighted anisotropic-isotropic total variation (AITV) has been applied successfully in various image processing problems, such as image denoising and image deblurring. For this presentation, we will investigate the application of AITV in variational image segmentation, creating new models and frameworks for accurate and effective segmentation.
Jae Hun Kim, Earth System Science – Automated recognition of grounding line migration from differential SAR interferometry using machine learning algorithms.
Abstract: Differential SAR interferometry (DInSAR) is a high precision space-based technique to detect transients in ground deformation for instance generated by the subtle vertical flapping of floating ice with changes in oceanic tides. DInSAR operates from 800 km away, at 50 m sample spacing, with millimeter precision in the vertical dimension, which is 10 to 100 times more precise than Geographic Positioning System (GPS) or laser altimetry, respectively. The ability to detect grounding lines, i.e. where a glacier becomes afloat in the ocean and detaches from the bed, has major implications for ice sheet/ocean modeling and projections of sea level rise from melting ice that will affect hundreds of millions of people worldwide. Since 2016, DInSAR observations of grounding lines have been acquired in vast quantities, at a 6-day repeat interval, around the coast of Antarctica, for years to come by ESA’s Copernicus program, with added capability down the line from CSA’s Radarsat constellation, ASI’s CSK constellation, JAXA’s ALOS-4 PALSAR-3, and NASA / ISRO’s NISAR. At present, DInSAR data are interpreted manually to delineate grounding lines because of the spatial complexities of the signal and the presence of data noise. Here, we propose to develop and evaluate an automated detection of grounding lines using machine learning algorithms that will detect the tidal migration of grounding lines and also reports on unusual variability, for instance due to a long term retreat of the grounding line consecutive to a de-stabilization of a glacier. The methodology will track down hundreds of glaciers simultaneously, on a 6-day repeat basis, all over Antarctica, for years to come. We will document the robustness of detection with respect to data noise and spatial variabilities in different settings, e.g. contrasting a smooth bed with a rough bed. We will compare the results with manual detection to assess reliability and error rates. We will assess the scalability of the approach with multi-sensor data and large data sets. This Artificial Intelligence (AI) approach should build the foundation for other detections of ground motion, e.g. subsidence of land associated with ground water withdrawal or pumping, caving of underground by human activities or natural processes, mineral or oil extraction on land, etc. that require time series of DInSAR and intelligent, scalable techniques of automated detection and analysis.