High entropy soft magnetic alloys
Faculty: Ian Baker
Location: MacLean ESC
Soft magnetic materials are used in a wide range of applications, from hair dryers and vacuum cleaners to electric vehicles and the aerospace industry. These materials can easily adjust their magnetization in response to a magnetic field, making them excellent material choices for power electronics. As our energy technologies continue to be modernized, developing novel soft magnetic materials that can keep up with these ever-evolving technologies becomes of great importance. Additive manufacturing (AM) offers an opportunity to easily test and optimize different novel alloy compositions and design structures. In this project, we will use laser powder bed fusion (a common AM method) to produce Fe40Co30Mn15Al15, an alloy that has excellent soft magnetic properties. We will investigate the printed microstructure and magnetic properties using a variety of different materials characterization techniques.
Game-theoretic reinforcement learning for cybersecurity
Faculty: Peter Chin
Location: ECSC
It is crucial for computer networks to be secure, trustworthy, and reliable and to respond to or even predict ever-increasing cyber-attacks. Advancements in deep learning, especially reinforcement learning (RL), offer hope to keep up with this challenge. However, the application of RL to cybersecurity has not achieved its full potential. Cyber-defense teams often fail to understand that their actions to defend the network are exactly what their adversaries are expecting. An adaptive and decentralized RL framework will be needed to combat the sophistication level of today's cyber-attacks. The student will work in the LISP (Learning, Intelligence + Signal Processing) Lab and help develop adaptive, hierarchical game-theoretic training frameworks that can predict future cyber-attacks.
Making federated learning more robust
Faculty: Peter Chin
Location: ECSC
Distributed learning paradigms such as federated learning often involve transmission of model updates, or gradients, over a network to avoid transmission of private data. However, it is possible for sensitive information about the training data to be revealed from such gradients. We have developed a method to discover the set of labels of training samples from only the gradient of the last layer and the id to label mapping. Our method is applicable to a wide variety of model architectures across multiple domains. The project will be to improve the effectiveness of our method for model training in various domains—image classification, automatic speech recognition, and more.
Deep routing in the real world
Faculty: Peter Chin
Location: ECSC
Routing through networks is a fundamental part of our lives, whether the route we take to a new restaurant or the data routed to us through the internet. For both kinds of networks, all routing approaches rely on the ability to approximate the distance between a node in the network and the target node, which can be accomplished by embedding the network into some metric space. Road networks are inherently embedded onto the surface of the earth, but there is no natural embedding for artificially constructed networks like the internet. It was observed, however, that routing on internet topologies works well when embedded into the hyperbolic plane. We want to explore how deep learning can help generate or adjust embeddings, to improve the performance of routing algorithms. To start, we will evaluate existing deep-learned embeddings for routing on a variety of real-world networks, and use the resulting insights to adapt and improve the embedding method.
Applying manifold learning technique to design recurrent architecture for low dimension classification
Faculty: Peter Chin
Location: ECSC
Deep Neural Networks (DNNs) can have high performance in a visual recognition task but are prone to noise and adversarial attacks. One main problem of training a DNN is that the input often lies in a high-dimensional space, which leads to a high number of parameters to train. This raises the question of reducing the number of dimensions of the dataset. Given a high-dimension dataset such as a visual dataset, how can we find a lower-dimension representation that keeps the essential information of the images? With a low-dimension representation, we can hopefully use a more shallow/simple architecture that can decently classify high-dimensional datasets.
Complex valued neural networks
Faculty: Peter Chin
Location: ECSC
Current neural network models that deal with data on the spectral plane (magnitude and phase) only take the magnitude as input, and do not incorporate the phase information in a meaningful way. Research has shown that the output of biological neurons is affected by the phase of its inputs. In order to bridge this separation between artificial neurons and biological neurons, I am experimenting with the effectiveness and implementation of complex-valued neural networks, which would integrate the phase information meaningfully. In particular, in order to take advantage of the popular neural network software package and framework, PyTorch, I am working to simulate complex-valued neural network operations through real-valued neural networks. I hope that by implementing complex valued neural networks through this framework, it will be easy for other researchers to use and experiment on.
Information propagation through graph neural networks and relation to the brain
Faculty: Peter Chin
Location: ECSC
A popular theory of intelligence argues that intelligence arises from the connections between primitive computing units rather than the computing units themselves. Interestingly, neurons in the brain form topological structures for processing different types of information. We are exploring the relationship between graph topology and model performance using graph neural networks, and comparing our findings to known phenomena in the brain.
Target propagation instead of back-propagation?
Faculty: Peter Chin
Location: ECSC
Deep neural networks trained with back-propagation have been the driving force for progress in fields such as computer vision and natural language processing. However, back-propagation has often been criticized for its biological implausibility. More biologically plausible alternatives to back-propagation, such as target propagation and feedback alignment, have been proposed. But most of these learning algorithms are originally designed and tested for feedforward networks, and their ability for training recurrent networks and arbitrary computation graphs is not fully studied nor understood. In this project, we propose a learning procedure based on target propagation for training multi-output recurrent networks. It opens doors to extending such biologically plausible models as general learning algorithms for arbitrary graphs.
IQ test for ML models
Faculty: Peter Chin
Location: ECSC
Current machine learning algorithms are highly specialized to whatever it is they are meant to do—e.g. playing chess, picking up objects, or object recognition. How can we extend this to a system that could solve a wide range of problems? We argue that this can be achieved by a modular system—one that can adapt to solving different problems by changing only the modules chosen and the order in which those modules are applied to the problem. The recently-introduced abstraction and reasoning corpus (ARC) dataset serves as an excellent test of abstract reasoning. Suited to the modular approach, the tasks depend on a set of human core knowledge inbuilt priors. In this project, we implement these priors as the modules of our system. We combine these modules using a neural-guided program synthesis.
Improving audio processing efficiency with deep learning and signal processing
Faculty: Peter Chin
Location: ECSC
Deep learning has significantly advanced the field of audio processing, transforming how audio data is analyzed, synthesized, and understood. The progress is largely due to deep neural networks (DNNs) and their ability to learn complex patterns from vast amounts of data, leading to innovations in various audio-related applications. However, deep learning has certain drawbacks, such as high computation/inference cost, high latency, privacy in training, etc. On the other hand, various techniques in classical signal processing are applied to many long-lasting problems in audio. For example, short-time Fourier transform (STFT) and matrix factorization are common methods applied to tasks such as denoising, cancellation, etc. The advantage of the classical signal processing is its low computation cost and low latency. Therefore, our focus is on how to improve audio processing efficiency by using the advantages of both deep learning and signal processing approaches. We focus on several audio signal processing tasks, namely low latency speech enhancement (and echo cancellation), privacy in training automatic speech enhancement, and audio codec.
Learning topological features via path complexes
Faculty: Peter Chin
Location: ECSC
Graph neural networks (GNNs), despite achieving remarkable performance across different tasks, are theoretically bounded by the 1-Weisfeiler-Lehman test, resulting in limitations in terms of graph expressivity. Even though prior works on topological higher-order GNNs overcome that boundary, these models often depend on assumptions about substructures of graphs. Specifically, topological GNNs leverage the prevalence of cliques, cycles, and rings to enhance the message-passing procedure. Our study presents a novel perspective by focusing on simple paths within graphs during the topological message-passing process, thus liberating the model from restrictive inductive biases. We prove that by lifting graphs to path complexes, our model can generalize the existing works on topology while inheriting several theoretical results on simplicial complexes and regular cell complexes. Without making prior assumptions about graph sub-structures, our method outperforms earlier works in other topological domains and achieves state-of-the-art results on various benchmarks.
Towards biopsy-free cervical precancer detection
Faculty: Irene Georgakoudi
Mentor: Matthew Lindley & Nima Najafi Ghalehlou
Location: Mostly remote
The overall goal of this project is to change the paradigm of cervical precancer detection (based on low-magnification visualization of the cervix and biopsy) using a non-invasive, high-resolution imaging modality that will ultimately provide valuable morphological and functional information at the bedside. We have collected such images from freshly excised human biopsies. Students will explore the use of an LLM-based image analysis platform (omega/napari) to denoise the images and segment cells so that their metabolic and morphological features can be recorded and used to establish algorithms that enable us to differentiate precancerous lesions from benign tissues and motivate such in vivo imaging in patients. Students will become familiar with image analysis as guided by AI algorithms and will gain an understanding of the pathophysiology of early cancers and the challenges that the medical community faces for their detection.
Developing a graphical user interface (GUI) for ex vivo tissue data collection
Faculty: Ryan Halter
Recommended background: Coding (MATLAB preferred)
Location: ECSC
We are investigating using impedance to assess cancer presence at surgical margins during lung wedge resections. To do this, we are collecting impedance data from ex vivo lung tissue with a trans-impedance measurement system (TIMS) that includes a height gauge and scale. For this study, it is important we collect data as soon as we can after the tissue is surgically removed. We are currently using a MATLAB script that requires users to push buttons on the height gauge, scale, and impedance analyzer. The goal of this project is to develop a graphical user interface (GUI) in MATLAB to help data collection go as smoothly and as quickly as possible. Ideally, the GUI will collect height and scale data with a single button push. Future versions would involve integrating the impedance analyzer into the GUI to allow for a seamless, unified interface. Opportunities for other coding and app development projects available but not required. Other opportunities to engage in data collection/experiments exist based on interest and availability.
Modeling thin film solar cells
Faculty: Geoffroy Hautier
Mentor: Zhenkun Yuan
Location: MacLean ESC
Solar energy, especially using the photovoltaic effect, is a critical component of a more sustainable energy system. The discovery of new materials for photovoltaic, especially thin-film photovoltaic, will be essential to a wider deployment of solar energy. Our research group has been running a project funded by the Department of Energy targeting the discovery of new materials for thin film photovoltaics. We have identified a few new exciting candidates for which we have modeling as well as experimental results. It remains to be known how these promising materials would work in a real solar cell device. This can be assessed by device modeling. The project will focus on learning how to use a standard device modeling software package using materials parameters for the emerging new materials discovered in our research group to model the efficiency of solar cell devices. The work will offer a clearer idea of the potential of these new materials as well as suggest specific device architecture.
Biomimetic combination construct for transcutaneous prosthetics support
Faculty: Katherine Hixon
Mentor: Adelaide Cagle
Location: ECSC
The project is targeted at developing tissue engineering scaffolds (composed of electrospinning and chitosan gelatin cryogels paired with 3D printing) for the support of medical implants that cross the dermal barrier. The research is specifically focused on improving the dermal seal around devices such as osseointegrated prosthetics, which exhibit poor adhesion to the surrounding skin. The main focuses of study are on the use of biomimetic simulation using keratin additives to induce dermal adhesion similar to nail bed adhesion to fingernails.
Swarm robot design, development, and optimization
Faculty: Michael Kokko
Location: Thayer
Students in ENGG 415: Distributed Computing will be remotely developing code to control a small swarm of robots with the intent of collectively mapping a 10ft x 10ft maze. We are developing a robotic platform for the first offering in 25W that is based on large, overpowered components, but we would ultimately like to shrink the footprint—likely developing a custom PCB and optimizing hardware for more nimble and reliable robots as the course scales to larger offerings. This FYREE opportunity will include learning from system use during the first offering in 25W and developing a second iteration of the hardware to be complete by the end of 25S.
AR app to aid in town planning and zoning
Faculty: Emily Monroe
Location: ECSC
The town of Hanover's board of Planning, Zoning, and Codes is responsible for planning Hanover's future in such areas as land use, economic development, transportation, natural resource protection, and maintenance and enhancement of Hanover's special character and quality of life. The board is looking to the future with more in-town housing options and needs help visualizing the impact of taller buildings on the look, feel, and sense of place in Hanover. This project proposes a mobile app using augmented reality to place proposed building designs over the top of current structures, with accurate shadows cast based on the time of day. This project will require front-end software engineering, an interest in urban planning, and a passion for more housing options in downtown Hanover.
Hanover GSHP feasibility
Faculty: Emily Monroe
Location: ECSC
In May 2017, Hanover residents voted overwhelmingly for the entire town to transition to 100% renewable electricity by 2030, followed by heating, cooling, and transportation by 2050. Inspired by Dartmouth's district heating geo-exchange project, Sustainable Hanover wishes to explore how ground-source heat pumps (GSHP) can help Hanover meet its 2050 thermal goal. The ideal outcome of this project is a feasibility study that looks at the town overall and suggests where geo-exchange heat pumps (or perhaps other renewable energy options) may or may not be appropriate, including for district heating. The project will consider urban and rural residential and/or small business properties, such as: those on Main Street; town-owned properties, including its schools; and large business properties, such as Kendal. This project will require expertise in energy engineering and mathematical modeling, a willingness to partner with local government and Hanover residents, and a passion for seeing change in the way energy is delivered in northern New England.
Improving usability of a portable iPhone-based method for EEG electrode localization
Faculty: Ethan Murphy
Location: Mostly remote; DHMC for occasional meetings
Our lab recently developed an iPhone-based approach to localizing electrodes for EEG. This is clinically important for source-localization techniques used in applications such as epilepsy, traumatic brain injury, psychiatric disorders, and brain tumors. Our implementation uses a GUI, which is kind of slow and clunky, in MATLAB. There are better software platforms, such as Python, to implement the GUI, and indeed, there are examples of similar implemented methods. The goal of this project is to transfer our current technique over to Python to improve the usability and speed of the localization process. You'll gain exciting programming experience in Python and potentially improve our tool for iPhone-based electrode localization.
Nanomedicine platforms for drug delivery and molecular imaging
Faculty: Hung Nguyen
Location: Rm 135, ECSC
Nanomedicine holds great promise in advancing modern medicine, exemplified by the rapid development of the mRNA vaccines that propelled the world through the COVID-19 pandemic. The modularity of this technology enables these vaccines to continue outpacing subsequent mutations. However, translation of this paradigm-shifting concept to other areas of medicine, including cancer therapy, remain challenging. This is mainly due to 1) the heterogeneity of cancer and 2) the complexities of treatment regimens. Towards this goal, we seek to develop convergent platforms for drug delivery and molecular imaging applications. Our initial focus includes combination therapy and cancer vaccines. Students will gain expertise in a broad spectrum of areas spanning chemical synthesis, material characterization, and biological examinations.
Genetic analysis of engineered strains of Clostridium thermocellum to improve cellulosic biofuel production
Faculty: Daniel Olson
Mentor: Ty Lanahan
Location: ECSC
Clostridium thermocellum is a bacterium that has a strong natural ability to consume cellulose. We are interested in harnessing this ability for cellulosic biofuel production. However, to do this, we need to improve the ability of this organism to produce ethanol, a potential biofuel. A bacterial cell is essentially a small chemical reactor that can be programmed to perform different chemical conversions. This is achieved by modifying the DNA sequences on the bacterial chromosome. However, techniques for chromosomal DNA modification often result in errors. Thus, a key part of the modification process involves checking the final DNA to make sure it has been modified as intended. The FYREE student will assist in the genetic analysis of strains of C. thermocellum to identify the presence of targeted genetic modifications (i.e., gene deletions or insertions). This genetic analysis will use classical molecular biology techniques of polymerase chain reaction (PCR) as well as more advanced techniques, including whole-genome resequencing.
In-vitro analysis of ethanol production pathway in Clostridium thermocellum
Faculty: Daniel Olson
Mentor: Shu Huang
Location: ECSC
Ethanol, produced from renewable sources like lignocellulosic biomass, is a promising biofuel that can help mitigate climate change by reducing reliance on fossil fuels. Clostridium thermocellum is a bacterium with significant potential for biofuel production due to its ability to efficiently break down lignocellulose. This project focuses on modifying the ethanol production pathway in C. thermocellum, aiming to maximize ethanol yield and titer. We will study proposed pathways in vitro using purified enzymes and measure their ethanol production levels. The findings will provide valuable insights and inspire strategies to optimize the pathway for improved performance in vivo. The FYREE student will assist in the expression and purification of these enzymes, gaining hands-on experience with essential lab techniques such as running protein gels, quantifying protein concentrations, and conducting enzyme assays to measure enzyme activities.
Detecting generated deceptive text from generative AI chatbots
Faculty: Eugene Santos Jr.
Recommended Background: This project will involve Python code development, experimental design, and statistical analyses. A sufficient background in coding is necessary. Experience in experimental design and statistical analyses is a plus. This project requires 10 hours of dedication per week.
Location: Rm 232, MacLean ESC
Large language models for generative AI produce text in tones and styles that essentially mimic anyone they have been exposed to in their training data, from the greatest speakers in history, the most compelling evangelists, and even the typical person on the street, to the most successful liars, shysters, and confidence tricksters. For the latter group, these models have ingested deceptive stories and writings, including both successful and unsuccessful ones. This project investigates the following questions: 1) Can we determine if a language model was trained on deceptive stories and writings? 2) Given a piece of text, can we come up with an algorithm to determine if it is deceptive in nature? This requires us to also define what deception is and what it is different from. 3) Generative AI hallucinates. How is this different or the same as deception?
Benchmarking low-power line-frequency transformers
Faculty: Charles R. Sullivan
Mentor: Allen Nguyen
Location: Rm 004, Cummings Hall
Class two line-frequency transformers are common for low-power applications such as HVAC systems, doorbells, and automatic doors. These transformers are heavy and bulky and waste large amounts of energy. As part of a larger project to develop more efficient technology for these applications, the student will perform electrical efficiency measurements across a wide selection (10–15) of these line-frequency transformers (LFTs) in order to generate performance benchmarks and to fully detail their power losses across their rated load range. Additionally, the student will design and build a two-layer printed circuit board (PCB) that streamlines this testing process. The student will learn the basics of: accurate power loss measurements; electrical safety; PCB design, build, and testing; and will understand the goals of modern switch-mode power electronics and power conversion.
Humidity impact on ferrite magnetic materials used in power electronics
Faculty: Charles R. Sullivan
Mentor: Thomas Guillod
Location: Rm 004, Cummings Hall
Nearly all high-performance modern energy technologies depend on power electronics and power electronics depends on high-frequency magnetic materials, primarily ferrites. Although ferrites' characteristics are known across a wide range of temperatures, the effect of humidity has not been studied much. The student will do systematic measurements to determine the effect of humidity, particularly on the permittivity of commercial ferrite materials. Tasks include preparing samples by depositing gold contacts, using humidity chambers, and making electrical measurements. The next step will be collecting and comparing the data, and analyzing the impact of humidity on their performance.
Algorithms for renewable energy transition in fragile states
Faculty: Vikrant Vaze
Mentor: Lilly Yang & Siqi Ke
Location: Remote
The worldwide transition to renewable energy and electrification is happening at an extremely uneven pace, with geopolitical stability being a key differentiator. Sixty percent of the 20 most energy-poor countries are in conflict regions, and many are in Africa. Diesel is the primary, and sometimes the only, source of energy in conflict regions. Burning of diesel emits greenhouse gases that exacerbate climate change and particulate matter that causes adverse health effects. Additionally, transporting diesel to far-flung remote locations presents severe costs and dangers. On the other hand, transitioning to renewables has tremendous peacebuilding potential. This project will build mathematical models and solution algorithms to support implementable decision-making frameworks for policymakers and other stakeholders to make a successful transition from diesel to renewable sources of energy in conflict areas.
AI for fighting wildfires
Faculty: Vikrant Vaze
Mentor: Spencer Bertsch
Recommended Background: Some knowledge in Python programming
Location: Remote
Climate change is creating warmer, drier conditions, leading to hotter and longer wildfire seasons, and effectively allocating firefighting resources such as trucks and planes is getting harder. We aim to aid firefighting decision-makers with distributed, deep multi-agent reinforcement learning methods that can help firefighters make better tactical decisions with the resources they have available. This problem becomes a dynamic resource allocation problem in which a constrained number of resources must be used to contain wildfires before they grow out of control. If you want to apply data science to fight climate change, would like to learn some hands-on AI skills, and have some background in Python programming, then this project is for you.
Data science for transportation planning and operations
Faculty: Vikrant Vaze
Mentor: Yueyun Xia
Location: Remote
Optimizing the design of large-scale transportation systems involves a large number of variables and constraints, and requires planners to balance objectives such as financial profitability, environmental sustainability, and equity. This project combines state-of-the-art methods in machine learning and optimization to tackle these notoriously hard problems.