Projects


01/2019 – Present, Postdoc, The Pennsylvania State University, University Park, PA, United States

Goals: Power of many – theory and development of domain independent deep learning solutions for Earth Science problems

Project-1: Develop deep learning solutions for volcanic surface movement monitoring (in Department of Geoscience, Penn State)

Project-2: Deep learning applications for automatic target detection and hyperspectral imaging with shortwave remote sensing data (collaborating with Department of Geography, Penn State)

  • Developed a temporal-dependent deep learning solution for atmospheric correction and target detection using multi-scan hyperspectral data

  • The time-dependency of the proposed network ensures the precision of estimated reflectivity spectra for different materials, including vegetations, sea ice, and ocean

  • Discussed the uncertainty and performances with missing data

  • Achievements:

    • Submitted to ISPRS Journal of Photogrammetry and Remote Sensing for peer-reviewing (under review)

    • Presented at AGU-2020 annual meeting

Project-3: Deep learning for sea ice concentration prediction and discovering the connections between sea ice and atmospheric factors (collaborating with Department of Atmospheric and Meteorology Science, Penn State)

  • Developed a neural network architecture to analyze the spatiotemporal variations of different atmospheric variables

  • Successfully predicted sea ice concentration using the proposed CNN-LSTM

  • Investigated the impacts of different atmospheric variables on sea ice concentration using interpretable deep learning

  • Achievements:

    • In Preparation: Deep learning discovery of predicting sea ice concentration in Sea of Okhotsk

Project-4: Hybrid deep learning for subsurface imaging and inversion (collaborating with Department of Geosciences, University of Calgary)

  • Developed a hybrid deep learning framework for seismic inversion

  • Discussed the advantages of physics-guided and data-driven deep learning

  • Comprehensively analyzed the deep learning performances for seismic inversion

  • Investigated the uncertainty of deep learning predictions of subsurface parameters

  • Achievements:

    • Submitted to Geophysics for peer-review: Sun et al., 2020

05/2018 – 01/2019, Postdoc, University of Calgary, AB, Canada

Project: Theory-guided deep learning in seismic forward and inversion problems

  • Designed a physics-based recurrent neural network (RNN) to simulate wave propagation in a deep learning perspective

  • Theoretically analyzed the deep learning formulation of seismic inversion problems

  • Proven that automatic differential technique is equivalent to the gradient-based full waveform inversion using adjoint-state methods

  • Implemented the seismic waveform inversion using the proposed theory-guided RNN

  • Discussed the hyperparameter selections for different optimizers and compared performances between auto-differential based algorithms and adjoint-state methods

  • Achievements:

    • Presented at CSEG-2019 and SEG-2019 annual meetings

    • Published a peer-reviewed article in Geophysics: Sun et al., 2020, which is the most downloaded article of all time in Geophysics Top-20

09/2013 – 05/2018, Research Assistant, University of Calgary, AB, Canada

Project: Data-driven theory for multiple prediction and attenuation in novel domains

  • Developed the inverse scattering series (ISS) theory for internal multiple prediction in the (coupled) plane-wave domain

  • Implemented multi-dimensional internal multiple prediction in the coupled plane wave domain with distinct advantages, such as constant hyperparameter selection, sparse implementation, and less computational overburden

  • Developed the ISS-based approach for multi-dimensional and multi-component internal multiple prediction in the (coupled) plane-wave domain

  • First successfully implemented the elastic internal multiple prediction in the multi-component seismic data

  • Achievements: