By Donsub Rim
During the academic year 2024–2025, I had the opportunity to return to the University of Washington (UW). My home institution, Washington University in St. Louis (WashU), grants its Arts and Sciences faculty an early sabbatical after their first three years. I decided to spend the year in Seattle: my wife was about to begin her job here, and I still had several active projects with my PhD advisor, Randy LeVeque.
It was more sentimental to be back in Lewis Hall than I had imagined. The tennis courts next to Lewis Hall were gone; the coffee shop in the basement of the art building was no longer there. The Ave had changed; Seattle overall had become busier, and the University District looked very different. Yet the important things had not changed. The sense of optimism and openness on campus and in Lewis Hall was still present. I ran into faculty members in Lewis and at Café Solstice and felt as if I were back in graduate school. Bernard was still Chair of Applied Math (though he has since stepped down after a long tenure). Many new faculty members had joined since my student days, but the department still felt familiar.
I was reminded that Ben Lansdell (a fellow UW applied math PhD) and I had served as system administrators during our graduate program. In my search for an external monitor, I met Ingrid Richter, the current system administrator, who told me she had seen our names in some of the documentation. I imagine our notes must have looked very amateurish in the eyes of a professional. I felt I could have done a better job had I known they would be read by others. But the encounter also reminded me of the great times I had working with Ben in Lewis.
The ongoing research projects with Randy also carried a sense of continuity, and I have worked on tsunami-related problems since graduate school, as it is a very rich topic. Together with Randy, collaborators Chris Liu and Bobby Baraldi (both UW applied math graduates), and a team of tsunami researchers led by Kenjiro Terada at Tohoku University, we are developing deep learning models for real-time tsunami prediction, with the potential to aid in early warning. Current warning systems cannot yet provide detailed, real-time predictions, and the hope is that new models will be able to forecast full tsunami waveforms and inundation patterns.
During a tsunamigenic earthquake, seismic waves travel much faster than tsunami waves. Thus, measurements of the earthquake become available within minutes, while the tsunami, even in near-field events, arrives tens of minutes later. To exploit this window of opportunity, we trained a neural network to predict the tsunami wave at a fixed nearshore gauge location over time, using geodetic measurements of the earthquake [1].
One challenge is the scarcity of real data, since large tsunamis are rare. Deep learning models rely on large datasets, so we turned to numerical simulations. Randy’s group at UW developed GeoClaw, a robust and efficient software package for simulating tsunami waves. Diego Melgar at the University of Oregon and collaborators contributed software for generating random earthquakes and efficiently propagating seismic waves through layered-earth models.
While our initial results were promising, we were keenly aware of a well-known shortcoming of deep learning models: their instability. A small perturbation in the input (for example, noise in earthquake measurements) can cause disproportionately large changes in the output (such as predicted tsunami wave height). Such instability, often referred to as adversarial examples, could have catastrophic consequences if predictions were wildly inaccurate. We are working to analyze and understand this phenomenon in order to stabilize the models [2]. Sanah Suri (a WashU graduate) and Tiana Johnson (a current WashU PhD student) are both involved in this effort.
I had the opportunity to participate in a Clawpack development workshop held at the University of Colorado Boulder, where I joined many researchers who were actively involved in developing the Clawpack software family. Many of them are UW applied math alumni, of course (see https://www.clawpack.org/geoclawdev-2025/). The venue was the Mountain Research Station in the beautiful mountains. Before this workshop, I gave a poster presentation at the CSDMS meeting. At the meeting, I enjoyed the company of Daniel Shapero (UW amath PhD), who gave an impressive Jupyter notebook clinic session on Icepack, a software package he has developed for modelling ice sheets. I recall seeing Daniel typing away excitedly in Café Solstice on numerous occasions during graduate school.
There were also numerous opportunities to present my research at UW. I spoke about Low Rank Neural Representations (LRNRs) at the Applied Mathematics Seminar and gave a talk on adversarial examples at the UW Data Science Seminar in Spring 2025. I also had the chance to share the tsunami project with K–12 students through the UW Women’s Center Summer Bridge Program, thanks to an invitation from Safi Karmy-Jones, another friend from graduate school. Together with Niket Thakkar (UW amath PhD), I talked with students about our journeys through graduate school and how they led to our careers.
It struck me that more than ten years have passed since I started graduate school at UW, and I have now known many of my peers for over a decade. When I wrote about my experience at UW Applied Math as I was graduating in 2017, I had declared that I had made lifelong friends. At the time I thought the phrase was a little hyperbolic. Now I am certain that, if anything, it was an understatement.
This week, we celebrate Randy’s 70th birthday. For a decade, I witnessed firsthand how his generous character imbued the UW community, and I remain deeply grateful to have been taught by him and to have shared in the UW applied math community.
References
[1] Rim, D., Baraldi, R., Liu, C. M., LeVeque, R. J., & Terada, K. (2022). Tsunami early warning from Global Navigation Satellite System data using convolutional neural networks. Geophysical Research Letters, 49, e2022GL099511. https://doi.org/10.1029/2022GL099511
[2] Rim, D., Suri, S., Hong, S., Lee, K., & LeVeque, R. J. (2024). A stability analysis of neural networks and its application to tsunami early warning. Journal of Geophysical Research: Machine Learning and Computation, 1, e2024JH000223. https://doi.org/10.1029/2024JH000223