2025 Research Prize Winners

1st Prize: Catherine Miller (Ecology, Evolution, Environment, and Society) Advisor: Jeremy DeSilva

This project investigates how early hominins walked by integrating 3D shape analysis of the distal femur with virtual musculoskeletal simulations. Using diffeomorphic shape matching, the study identified key morphological variations linked to knee flexion during gait. Fossil hominins generally exhibited more flexed postures than modern humans. To test the biomechanical feasibility of these postures, virtual musculoskeletal simulations of Australopithecus afarensis and a modern human were run in OpenSim. Results showed that moderate knee flexion was sustainable only with hip extension. These findings support a novel extended-hip-bent-knee model for bipedal gait in early hominins—distinct from both modern humans and chimpanzees.

2nd Prize: Ivory Yang (Computer Science) Advisor: Soroush Vosoughi

The preservation and revitalization of endangered languages, especially those with limited digital presence, poses considerable challenges for the field of computational linguistics. In response, we harness the power of Artificial Intelligence (AI) and Large Language Models (LLMs) to breathe new life into these languages, facilitating the creation of digital resources and models from scarce data.This work presents a compilation of three papers on Nüshu (published at COLING 2025: https://aclanthology.org/2025.coling-main.468/), Navajo (Native American) (published at NAACL 2025: https://aclanthology.org/2025.naacl-short.24/) and Native Alaskan languages (under review at ACL 2025).

2nd Prize: Juhyeon Kim (Computer Science) Advisor: Adithya Pediredla

Accurate digital twins of imaging systems are essential for computational imaging research. In this project, I introduce a Monte Carlo rendering framework for simulating the Doppler effect in velocity-sensitive imaging systems, contributing physically accurate simulation methods for two distinct but related sensing modalities: Doppler time-of-flight (D-ToF) cameras and optical heterodyne detection (OHD) systems.

Personal website : https://juhyeonkim.netlify.app, D-ToF (SIGGRAPH Asia 2023, ToG) : https://dl.acm.org/doi/10.1145/3618335, OHD (SIGGRAPH 2025, ToG) : appears in July

3rd Prize: Xiangbei Liu (Thayer School of Engineering) Advisor: Yan Li

Metamaterials with zero Poisson's ratio they maintain their shape in the transverse direction when stretched or compressed, making them essential for technologies like soft robotics and biomedical devices. However, designing such materials remains a major challenge due to the vast design space and the lack of existing samples. We present a few-shot machine learning framework that uses a novel conditional variational autoencoder to generate manufacturable and customizable designs with unprecedented efficiency. By combining active learning and simulation with experimental validation, our method boosts success design rates from 0.3% to 39% with limited and biased data, enabling rapid discovery of complex material architectures and opening new frontiers in multifunctional metamaterials and generative design.

Few-shot learning-based generative design of metamaterials with zero Poisson's ratio. Material & Design 2024:113224. https://doi.org/10.1016/j.matdes.2024.113224

3rd Prize: Mingi Jeong (Computer Science) Advisor: Alberto Quattrini Li

We present a best-in-class method called active learning-augmented intent-aware obstacle avoidance. Using a Long Short-Term Memory (LSTM) neural network and a novel topological modeling of passing behaviors, our approach infers the passing intentions of obstacles. The learning-augmented algorithm ensures both safety guarantees and interpretability. We demonstrate its effectiveness through a real marine accident case study and real-world experiments with an ASV operating under environmental disturbances, showing successful real-time collision avoidance. This method has strong potential for high-impact applications, including maritime transportation, environmental monitoring, and search and rescue. The work has been published at IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2024 https://ieeexplore.ieee.org/document/10802205

1st Prize: Amit Das (Computer Science, Biomedical Data Science, Epidemiology) Advisor: Saeed Hassanpour

My research introduces HistoStainAlign, a novel deep learning framework for predicting immunohistochemistry (IHC) staining patterns directly from hematoxylin and eosin (H&E) whole-slide images. By aligning morphological and molecular features using contrastive learning, this model reduces the need for costly and time-consuming IHC procedures. Evaluated on gastrointestinal and lung tissues for P53, PD-L1, and Ki-67 biomarkers, the framework demonstrated strong predictive accuracy and cross-modality alignment. This work has applications in digital pathology, enabling faster and more cost-effective cancer diagnostics.

2nd Prize: Karun Ram (Computer Science) Advisor: Siddhartha Jayanti

In this paper, we consider a near-optimal randomized multiprocessor algorithm for the union-find problem due to Jayanti and Tarjan, and provide the first machine-verified proof of its correctness. Our proof, written in TLA+ and verified using TLAPS, is intricate—spanning over 16,000 lines and organized into 20 theorems. Further, our modelling of the object, and the corresponding proof, introduce novel methods which could aid future machine-verification efforts. Because the proof is machine-certified, algorithm designers can rely on the correctness of these union-find implementations without manually inspecting the proof details, thereby mitigating the risk of catastrophic race conditions and concurrency bugs.

2nd Prize: Mingyue Zha (Quantitative Social Science) Advisor: Herbert Chang

Content-creator collaborations enhance digital viewership and revenue, yet few studies examine gender inequities in these dynamics. Analyzing 42,376 YouTube videos and 6.1 million comments across 150 channels and 3 games, this study uses Shapley value from cooperative game theory to reveal in-group collaboration patterns shaped by game affordances. Despite genre differences, audience biases remain consistent, reflecting symmetric biases across the gaming communities. Thus, we find supply-side gender asymmetries and demand-side symmetries. Our results build on digital bias literature, demonstrate how genre and affordances shape gendered collaboration and direction of inequality, and provide a framework to quantify synergy.

3rd Prize: Matthew Timofeev (Engineering) Advisor: William Scheideler

Perovskite solar cells are emerging alternatives to silicon due to their high efficiency and relative ease of fabrication, but their power stability when combining solar cells into large-scale modules remains a drawback. This project builds towards a design aid for fault-tolerant perovskite solar modules, based on extensive finite-element simulation and predicting solar module performance given expected distributions of defects within component cells, such as pinhole shunts and excessive series resistance. This exploits cases where poor shunt and series resistance may combine and preserve module performance, presenting an opportunity for rapid design iteration in novel solar module designs for stability.

https://pubs.aip.org/aip/ape/article/1/3/036104/2918603

3rd Prize: Pratim Chowdhary (Computer Science) Advisor: Peter Chin

This paper introduces K-MSHC, a framework for identifying minimal sets of attention heads crucial for specific tasks in language models. Using Gemma-9B, researchers discovered that different capabilities utilize distinct neural circuits: grammar tasks primarily engage early layers (0-6), arithmetic verification shows distributed patterns across the network, and word problems activate both shallow and deep regions.

Key findings: Tasks share "weak" computational components but maintain dedicated "super-heads" with minimal overlap for critical processing. This reveals that language models develop specialized yet partially reusable circuits rather than fully general mechanisms.

Importance: This work advances mechanistic interpretability by providing tools to understand how language models organize capabilities, potentially enabling targeted interventions to enhance specific abilities without disrupting others.