School of Science Department of Ocean Science 58 Mechanistic Interpretability of Machine Learning Models in the Physical Sciences Supervisor: Julian MAK / OCES Student: SO Hayden / IIM Course: UROP 1000, Summer Interpretability of neural networks in the physical sciences remains unexplored. In this paper, we present a novel approach to tropical cyclone (TC) intensity estimation through the application of a Sparse Feature Network (SFNet). Unlike traditional “black box” convolutional neural networks, the SFNet model constrains activations to a sparse set of interpretable features, which allows for a post-hoc transparent analysis of the visual patterns associated with tropical cyclone intensity. We demonstrate that SFNet achieves benchmark accuracy in estimating tropical cyclone wind speeds while providing mechanistic interpretability (MI) of the specific visual features used in each individual prediction. These interpretable features show alignment with known meteorological phenomena correlate with cyclone intensity. Our work bridges the gap between traditional expert-based methods and deep learning approaches, offering scientifically meaningful insights. Beyond, this demonstrates how MI research can benefit foundation models in the physical sciences, where understanding “why” and “how it learns” matters. Coral-Symbionts Interaction and Evolution Supervisor: WU Longjun / OCES Student: WONG Wing Ki / BCB Course: UROP 1100, Fall Oceanic warming is an increasingly pressing issue that is closely intertwined with the broader challenges of climate change. Zooplankton is a diverse group of heterotrophic organisms that inhabit aquatic ecosystems ranging from freshwater to marine environments. They inevitably have to cope with drastic heat content and thermal stress surges. Despite their microscopic size, they are the main secondary producers in oceans and significantly contribute to marine diversity through the nutrient loop (Hobday et al., 2006; Thackeray, 2022). However, their underlying mechanisms in the dissolved organic matter (DOM) cycle remain understudied, especially under ocean warming scenarios. To explore the composition changes of zooplankton-derived DOM, we will assess zooplankton-derived DOM incubation across three different temperature scenarios. The subsequent release DOM will be collected using the Horiba Scientific Aqualog® Fluorometer, and statistical analysis will be conducted. Our ultimate goal is to provide a more comprehensive understanding of zooplankton’s role in the marine DOM cycle. Exploring the Mechanisms of the Reverse Development of Immortal Jellyfish Turritopsis Supervisor: WU Longjun / OCES Student: HUI Wang Kit Zachary / BCB Course: UROP 1100, Fall This progress report outlines the research experience I gained in Dr. Wu’s lab from September to December. It includes my work on raising polyps of an undescribed species of hydrozoan. This jellyfish is capable of reverse development, transitioning from polyps to medusa and back again. Over the past three months, I have raised polyps of this species and assisted in conducting trials to determine the optimal conditions for their maintenance, including identifying suitable feeders, determining the optimal temperature, and addressing issues related to protists and biofilm. Additionally, I conducted PCR tests on this species. I accomplished all of this under the guidance of postgraduate student, Shen Lin and Dr. Wu.
RkJQdWJsaXNoZXIy NDk5Njg=