Intersections of Machine Learning and Physical Science: Three Use Cases

Jennifer Sleeman
  23 Oct, Noon, in google meet

In this talk, I will present my research in three areas that exhibit an intersection between machine learning and physical science. I will describe my thesis work for which I took the theory of data assimilation and applied it to natural language, processing over 30 years of the Intergovernmental Panel for Climate Change (IPCC) reports and their citations. I will show how we used this method to detect both relatedness among research papers and their influence on the IPCC reports. I will then describe my work related to developing a new hybrid method to overcome qubit limitations when training Restricted Boltzmann Machines on the D-wave quantum computer. We were able to show a 22-fold compression and our findings showed that we could use the natural inherent randomness from the quantum computer for image generation. Finally, I will describe my current work which uses a deep segmentation neural network to identify aerosol boundary layer heights (ABLH) for ceilometer-based Lidar backscatter profiles. We employ a novel machine learning method that uses near time-continuous profiles forming an image to determine boundary layer heights rather than wavelet covariance approaches to determine profile discontinuities that are often noisy or affected by clouds. We developed a method that works between pixel and coordinate spaces and were able to show our deep segmentation algorithm is able to detect heights that are validated and competitive with radiosonde traditional PBLH methods. In addition, I will describe the current efforts exploring the comparison of ABLHs derived from the WRF-Chem-GOCART model profiles with ceilometer profile ABLHs. I will also provide a preview of our effort to extend the current transfer learning edge detecting algorithm to a stacked Long Short Term Memory (LSTM) neural network to simultaneously process Lidar backscatter data from the Icesat-2 satellite, ceilometer backscatter, and cloud-resolving model generated aerosol Lidar backscatter. onstraint and Soil Moisture Active Passive (SMAP) for the Sea Surface Salinity (SSS) constraint.