OUTSTANDING Graduate Research in Computational Science
1st Prize: Dario Seyb (Computer Science) Advisor: Wojciech Jarosz
Detailed geometric models are necessary to faithfully simulate physical processes. But even with modern computing resources, we cannot represent details such as the individual rain drops in a cloud or the microscopic bumps on a rough surface explicitly. Instead, existing methods use statistical models which are often approximate and highly specialized to a certain type of geometry.
We present a unified statistical model of geometry using Gaussian process implicit surfaces. Our framework allows us to express a wide range of existing models and move between them in a continuous fashion, opening up both new theoretical insights as well as practical applications. Dario Seyb, Eugene d'Eon, Benedikt Bitterli, and Wojciech Jarosz. 2024.
From microfacets to participating media: A unified theory of light transport with stochastic geometry. ACM Transactions on Graphics (Proceedings of SIGGRAPH). 43, 4, Article 112 (July 2024), 17 pages. https://doi.org/10.1145/3658121
1st Prize: Alex Gottlieb (Ecology, Evolution, Environment and Society (EEES)) Advisor: Justin Mankin
Despite the widespread intuition that a warmer world is a less snowy one, identifying where, when, and by how much climate change has affected critical snow water resources has remained elusive. In this project, we leverage over a dozen observational datasets of snowpack, temperature, and precipitation and hundreds of global climate model simulations to show a clear human fingerprint of greenhouse gas emissions on spring snowpack and the vital runoff it generates in many highly-populated Northern Hemisphere river basins. Crucially, we also identify a "snow loss cliff", where locations that are warmer than -8°C during the winter experience accelerating snow loss with each degree of warming, portending sharp snow declines and water security challenges absent aggressive mitigation.
2nd Prize: Tommy Botch (Psychological & Brain Sciences) Advisor: Emily Finn
Large language models (LLMs) are increasingly used across disciplines and, within the field of cognitive psychology and neuroscience, are commonly suggested as substitutes for human behavior. However, humans learn language primarily through a spoken modality while LLMs are often trained solely on written text. Our work evaluates how stimulus poverty – here, the removal of auditory information – impacts language processing and prediction within humans and LLMs. We found that human predictions of spoken language were more accurate and more closely aligned with human brain activity than both human predictions of written text and LLM predictions. Together, these results suggest that human predictions of written text represent a theoretical ceiling on what LLMs can achieve in both behavior and representation.
3rd Prize: Quinton (Ziyuan) Qu (Computer Science) Advisor: Adithya Pediredla
Differentiable 3D-Gaussian splatting (GS) is emerging as a prominent technique in reconstructing 3D scenes. GS represents a scene as 3D Gaussians with analytical derivatives to compute their parameters from images taken from various viewpoints. Unfortunately, capturing surround view images is impractical in many real-world imaging scenarios, including underwater imaging, rooms inside a building, and autonomous navigation. In these scenarios, the GS algorithm suffers from 'missing cone' problem, which results in poor reconstruction along the depth axis. We demonstrate that using transient data allows us to address the missing cone problem by sampling high-frequency data along the depth axis. We extend the Gaussian splatting algorithms for two commonly used sonars and propose fusion algorithms that simultaneously utilize RGB camera data and sonar data.
3rd Prize: Abdullah Al Maruf (Earth Science) Advisor: Sarah Slotznick
The presence of water on extraterrestrial Earth-like bodies has captivated the scientific community for its implications on habitability. Hematite (α-Fe2O3), an iron-oxide mineral prevalent on Mars, has been investigated as a potential carrier of structurally-bound water, transforming into a unique phase called hydrohematite (hyhm), (Fe2−x/3O3−x(OH)x), in nature. Our atomistic simulations from first-principles approach demonstrate that hematite can incorporate up to ~20% structural water (–OH) per unit cell by substituting Fe atoms, resulting in up to 17% Fe vacancies. Synchrotron-XRD experiments corroborate these findings, showing approximately 16% Fe vacancies in natural hyhm and around 12% in synthetic samples. Both computational and experimental results indicate significant changes in the magnetic properties in water containing hematite, particularly the complete suppression of the Morin transition (TM) at low temperatures – proving hym's existence in nature. The results contribute to our understanding of potential water-signature in hematite in arid-planetary environments and can serve as a basis for non-destructive magnetic method to advance the ongoing search for water on Mars.
Honorable Mention: Rayna Rampalli (Physics and Astronomy) Advisor: Elisabeth Newton
It has been shown that the Sun appears relatively depleted in certain elements compared to other sun-like stars in the Galaxy. This depletion could be attributed to certain planet formation processes within our solar system. We test if this is true by inferring elemental abundances for over 17,000 Sun-like stars, some of which are known planet hosts, using a data-driven model. We find that the Sun remains chemically anomalous compared to other stars that host similar sized planets. This suggests that various planet formation mechanisms alone do not account for the Sun's unique chemistry; other solar system or galactic processes could be at play with potential implications for constraining habitability signatures.
Honorable Mention: Xie He Mathematics Advisor: Peter Mucha
Link prediction algorithms are indispensable tools in many scientific applications by speeding up network data collection and imputing missing connections. However, in many systems, links change over time and it remains unclear how to optimally exploit such temporal information for link predictions in such networks. Here, we show that many temporal topological features, in addition to having high computational cost, are less accurate in temporal link prediction than sequentially stacked static network features. This sequential stacking link prediction method uses 41 static network features that avoid detailed feature engineering choices and is capable of learning a highly accurate predictive distribution of future connections from historical data. We demonstrate that this algorithm works well for both partially observed and completely unobserved target layers, and on two temporal stochastic block models achieves near-oracle-level performance when combined with other single predictor methods as an ensemble learning method. Finally, we empirically illustrate that stacking multiple predictive methods together further improves performance on 19 real-world temporal networks from different domains.