31825 research outputs found
Sort by
Optimizing and scaling machine learning models for scientific applications on exascale supercomputers
No full textChandrasekaran, SunitaAs software and hardware concurrently advance, it's important to design and build ML frameworks that are hardware architecture agnostic. More specifically as accelerators for ML workflows become more prevalent, the ability to have high level code that can run across such accelerators would be highly beneficial by reducing the need to rewrite code and libraries for each hardware. At the same time, advancement in machine learning (ML) methods has enabled extraction of meaningful information from large and complex datasets that have assisted in better understanding, diagnosing, and treating illnesses such as cancer. This applies to applications beyond oncological drug response and drug discovery including understanding complex plasma physics phenomena. ☐ This thesis focuses on designing and building scalable and portable machine learning-based workflows while adapting them to new hardware architectures. This thesis also includes scaling and improving performance of surrogate models that reduce scientific simulations necessary for extracting insights. A phenomenon increasingly necessary as scientific challenges increase in computational complexity. We demonstrate the ideas using two case studies. ☐ An improved drug discovery pipeline is designed for shorter development timelines through model enhancement and scaling on new hardware capabilities. The thesis investigates gradient boosted tree-based methods as viable alternatives to CNNs in demonstrating the limitations of existing neural network-based drug response models. These gaps are resolved by designing and building software that helps assess the variation in performance of each class of models and includes improvements made to the accessibility of these models for domain experts. The current approaches rely on RNA sequence based gene expression values of cell lines, 2D molecular drug descriptors, and drug response data to predict cell growth. To overcome the challenges faced with the existing 2D molecular datasets, the next aspect of the thesis focuses on improving the performance of ML techniques that synthesize molecular docking of the 3D molecular drug descriptors for pose estimation of protein-ligand binding to reduce the subsequent molecular dynamics simulations needed in drug-discovery workflows. In addition to hyperparameter optimization (HPO) and model tuning, scaling the training of such models will greatly improve the throughput of lead compound discovery. ☐ Scaling such ML workflows on new hardware architectures like AMD GPUs is challenging. The thesis further explores the scaling aspect by using another case study that involves in-transit ML of plasma physics simulations to uncover correlations between emitted radiation and particle dynamics within the simulation. The ML surrogate model employs online learning using data streamed from the simulation and scales up to 400 GPUs. ☐ To summarize, this thesis introduces novel software frameworks and workflows to advance the state-of-the art for case studies involving drug discovery models for cancer research as well as plasma physics simulations through model enhancement and distributed scaling on large supercomputers.University of Delaware, Department of Computer and Information SciencesPh.D
Electrochemical deposition of vanadium and manganese oxides for use as positive electrodes in rechargable batteries
No full textDavidson, Rachel A.Energy storage technologies represent an essential component in redesigning the electric grid to mitigate CO2 emissions and mitigate climate change. Despite the substantial progress which has been made in recent years in electrification of the transportation industry, battery performance metrics such as the charging rate and time have hindered growth. Charging rate is closely linked to battery stability. For a given battery, charging at a faster rate requires greater overpotential which provides a greater thermodynamic driving force for reaction and promotes out of equilibrium reactions. Therefore, improvements in the stability of battery systems can also have implications for the charging rates which can be applied. The theoretical limits on battery capacity are largely tied to the material composition and crystallographic structure of electrodes, which for the positive electrode (cathodes) are related to the number of intercalation sites versus the weight and volume of the material. However, in practice the availability of these active sites during initial charge and their availability over time is substantially affected by issues in transport, crystal defects and transformations, and the stability of the electrode over time. Therefore, controlling stability and minimizing limitations in transport which hinder kinetics of charge/discharge represent crucial challenges in battery design.University of Delaware, Department of Chemistry and BiochemistryM.S
Improving foundation models on electronic health records
No full textBeheshti, RahmatollahRecent advances in foundation models have opened up new possibilities for healthcare applications, particularly by utilizing transformer-based models to take advantage of the longitudinal nature of both natural language and electronic health records (EHRs). While these models have shown promise, existing approaches face challenges related to multi-task learning; knowledge transfer from pre-training to finetuning stages; simultaneous representation of medical codes and visits; and potential social biases in predictions. ☐ The primary goal of this dissertation is to tackle these issues by presenting multiple transformer-based models while investigating and mitigating their issues related to fairness. Our proposed solutions have been evaluated on a multitude of popular medical predictive tasks. We first propose a transformer-based model tailored for multi-task learning, used for the primordial prevention of cardiovascular disease. Second, we tackle the issue of decreasing performance on small datasets with a semi-supervised transformer model that leverages both in- and out-of-cohort patients in the context of few-shot learning. Third, we propose a hybrid model that leverages graph neural networks to extract the structure of medical visits, and a transformer encoder to extract the temporal relationships of visits. Fourth, we investigate the fairness implications of our models and propose a bias mitigation technique based on federated learning principles. Lastly, we investigate the specific challenges of fairness in medical large language models (LLMs), conducting a comprehensive evaluation of the bias patterns. We then present a novel bias mitigation technique for medical LLMs based on model alignment ideas within a knowledge distillation framework.University of Delaware, Department of Computer and Information SciencesPh.D
Assessment of Graduated Compression Stockings: Variation in Fabric Properties Among Different Sizes
No full textThis article was originally published in International Textile and Apparel Association Annual Conference Proceedings. The version of record is available at: https://doi.org/10.31274/itaa.18824.
© 2024 The author(s). Published under a Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly citedThe purpose of this study was to investigate the physical properties of one size range (S-XXL) of a commercial brand of medical grade GCS and determine how the material properties change between sock sizes. Understanding these differences could help GCS manufacturers improve their sizing design process, resulting in more consistent pressure properties across sock sizes. One pair of a complete size range of knee-length medical grade GCS were purchased from a reputed domestic manufacturer. The medium compression GCS were made of 90%bamboo 10%spandex, with a pressure range of 20-30mmHg, knitted on Lonati 240N circular sock machine. ANOVA was used to evaluate the effect of ‘SockSize’ and Level (‘Ankle’ and ‘Calf’) on all dependent variables. The results of this study highlight the inconsistent fabric properties across a size range of medical grade GCS, with the larger sizes having potentially lower pressures on the leg than the smaller sizes
Comparative metagenomics of tropical reef fishes show conserved core gut functions across hosts and diets with diet-related functional gene enrichments
No full textThis article was originally published in Applied and Environmental Microbiology. The version of record is available at: https://doi.org/10.1128/aem.02229-24.
Copyright © 2025 Wu et al. This is an open-access article distributed under the terms of the Creative Commons Attribution 4.0 International license (https://creativecommons.org/licenses/by/4.0/).Fish gut microbial communities are important for the breakdown and energy harvesting of the host diet. Microbes within the fish gut are selected by environmental and evolutionary factors. To understand how fish gut microbial communities are shaped by diet, three tropical fish species (hawkfish, Paracirrhites arcatus; yellow tang, Zebrasoma flavescens; and triggerfish, Rhinecanthus aculeatus) were fed piscivorous (fish meal pellets), herbivorous (seaweed), and invertivorous (shrimp) diets, respectively. From fecal samples, a total of 43 metagenome assembled genomes (MAGs) were recovered from all fish diet treatments. Each host-diet treatment harbored distinct microbial communities based on taxonomy, with Proteobacteria, Bacteroidota, and Firmicutes being the most represented. Based on their metagenomes, MAGs from all three host-diet treatments demonstrated a baseline ability to degrade proteinaceous, fatty acid, and simple carbohydrate inputs and carry out central carbon metabolism, lactate and formate fermentation, acetogenesis, nitrate respiration, and B vitamin synthesis. The herbivorous yellow tang harbored more functionally diverse MAGs with some complex polysaccharide degradation specialists, while the piscivorous hawkfish’s MAGs were more specialized for the degradation of proteins. The invertivorous triggerfish’s gut MAGs lacked many carbohydrate-degrading capabilities, resulting in them being more specialized and functionally uniform. Across all treatments, several MAGs were able to participate in only individual steps of the degradation of complex polysaccharides, suggestive of microbial community networks that degrade complex inputs.
IMPORTANCE
The benefits of healthy microbiomes for vertebrate hosts include the breakdown of food into more readily usable forms and production of essential vitamins from their host's diet. Compositions of microbial communities in the guts of fish in response to diet have been studied, but there is a lack of a comprehensive understanding of the genome-based metabolic capabilities of specific microbes and how they support their hosts. Therefore, we assembled genomes of several gut microbes collected from the feces of three fish species that were being fed different diets to illustrate how individual microbes can carry out specific steps in the degradation and energy utilization of various food inputs and support their host. We found evidence that fish gut microbial communities share several core functions despite differences in microbial taxonomy. Herbivorous fish harbored a functionally diverse microbial community with plant matter degraders, while the piscivorous and invertivorous fish had microbiomes more specialized in protein degradation.We thank our colleagues at UD in Lewes for assistance with fish care and sampling.
This work was supported by The Mindlin Foundation, funding to C.R.H. via the UD Graduate Scholars Fellowship, funding to K.M.K. and J.F.B. from the WM Keck Foundation, and funding to I.F.F. and J.F.B. by ExxonMobil Research and Engineering Company. D.G.W. was supported as a UD School of Marine Science & Policy scholar. Support from the University of Delaware CBCB Bioinformatics Data Science Core Facility (RRID:SCR_017696), including use of the BIOMIX and BioStore computational resources, was made possible through funding from Delaware INBRE (NIGMS P20GM103446), NIH Shared Instrumentation Grant (S10OD028725) the State of Delaware, and the Delaware Biotechnology Institute
Promoting Students’ Engagement in Civil Dialogue: A Pilot Study and Randomized Controlled Trial
No full textIn this report, I describe the results of a pilot study and a randomized controlled trial that investigated the possibility that engaging in a structured conversation with someone of differing political views can increase students’ willingness to express their views on controversial topics and improve students’ attitudes toward individuals who do not share their views.
The pilot study (n=47) was conducted during the fall 2023 semester. Students from the University of Wisconsin-Eau Claire (UWEC) were asked to participate in a Unify America College Bowl; a virtual one-on-one conversation with a student at a different university who holds different political views . Afterward, students reported on their experience through an online survey. As a whole, the students reported enjoying their conversations very much and rated their conversation partner and themselves as acting very respectfully. They reported that they began the conversation feeling quite nervous and that they ended the conversation quite optimistic and inspired. In addition, more than half of the students who participated reported that, after the conversation, they placed more value on viewpoints that differ from their own and felt more comfortable interacting with people with opposing views.
However, the students in the pilot study were a select group of students: directly or indirectly, they had chosen to participate in the virtual conversation. They may have been particularly interested in engaging in civil dialogue, open to hearing diverse viewpoints, or willing to take the psychological risk of meeting a stranger online. Perhaps the positive trends in the pilot study had less to do with engaging in a civil dialogue than with the type of students who choose to engage in civil dialogue.
Thus, a randomized controlled trial was conducted during the spring 2024 semester. To launch the study, an email from the chancellor encouraged continuing first-year students to participate in a semester-long study of “constructive conversations.” Participating students reported their attitudes about viewpoint expression and viewpoint diversity at the beginning (Phase 1), middle (Phase 2), and end of the semester (Phase 3). By random assignment, students in the “conversation” condition were asked to engage in a one-hour civil dialogue through Unify America College Bowl just before responding to the Phase 2 survey; students in the control condition were asked to watch a one-hour neutral video before responding to the Phase 2 survey.
For the 50 students in the conversation condition who completed all three phases, there were key changes that aligned with expectations. First, their perception of how much disagreement exists between them and other college students of differing viewpoints decreased from baseline to immediately after the College Bowl, and their perception of how much agreement exists increased. Second, the degree to which they felt favorably toward students of differing viewpoints increased after engaging in the College Bowl, and this change was maintained in the Phase 3 follow-up. For the 116 students in the control condition who completed all three phases of the study, changes in attitudes over the three phases were inconsistent.
Notably, more students in the conversation condition than in the control condition did not complete the online task, essentially dropping out of the study during Phase 2. During Phase 3, we learned that students in the conversation condition who did not proceed with the College Bowl most often reported it was because they were uncomfortable speaking with a stranger online (64%), didn’t have time (52%), or weren’t interested in engaging in an online conversation about political issues (44%). Students who did engage in the conversation said it was because they wanted to earn the financial incentive (92%) and they were curious about what the online conversation would be like (70%). The open-ended responses from students who did engage in the conversation were nearly all positive (mean rating of 9 on a 0-10 scale).
Together, the findings imply that many students, perhaps mostly out of fear/anxiety or disinterest, will probably not actively engage in civil dialogue unless they are required to. However, among students who do actively choose to engage in civil dialogue, positive attitudinal change can occur, at least in the short-term. To promote both attitude change and behavioral change, future initiatives might expose all students to civil dialogue (e.g., through Unify America) and follow it with continued opportunities for civil dialogue and intentional, explicit discussion of personal characteristics (e.g., openness) and behaviors (e.g., asking questions, sharing personal stories) that promote constructive conversations.SNF Ithaca Initiativ
Cropland expansion links climate extremes and diets in Nigeria
No full textThis article was originally published in Science Advances. The version of record is available at: https://doi.org/10.1126/sciadv.ado5541.
Copyright © 2025 The Authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original U.S. Government Works. Distributed under a Creative Commons Attribution License 4.0 (CC BY).
This is an open-access article distributed under the terms of the Creative Commons Attribution license (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
This research was featured in UDaily on 1/15/2025 at: https://www.udel.edu/udaily/2025/january/nigeria-agriculture-deforestation-farming-farms-food-cropland-expansion-climate-change/Climate change threatens smallholder agriculture and food security in the Global South. While cropland expansion is often used to counter adverse climate effects despite ecological trade-offs, the benefits for diets and nutrition remain unclear. This study quantitatively examines relationships between climate anomalies, forest loss from cropland expansion, and dietary outcomes in Nigeria, Africa’s most populous country. Combining high-resolution data on forest cover and climate variables within random forest and panel regression models, we find that 25 to 31% of annual forest loss is linked to climate variability. Using georeferenced household survey data, we then find that changes in forest cover have a significant positive association with changes in child diet diversity—a key proxy of nutritional adequacy—while cropland expansion does not, suggesting that such forest conversions may be an ineffective climate adaptation strategy for improving nutrition. Our findings highlight the potential of nutrition-sensitive climate adaptation to enhance yields, promote nutritious cropping choices, and protect remaining forests.We acknowledge that we received no funding in support for this research
