Projects

Informatics Skunkworks members are working on a range of research projects which are summarized in the table below. Prospective participants are encouraged to reach out to the listed team leads to inquire about participation in their respective projects.

In addition to projects hosted by mentors at UW, there are a number of projects hosted by collaborators at BSU and students are able to participate remotely if they have strong interest in a project at another institution. Please see the Boise projects page here

More detailed information on the projects below can be found in the full project spreadsheet

Skunkworks Projects

Short Name	Title	Project Description	Application deadline	Requirements: hrs/wk	Requirements:skills/courses/other	Contact: First Name	Contact: Last Name	Contact: Department	Contact: Email	Contact: Phone	Team leads (Names, Departments, Emails)	Number of Positions (max)	Funds available?	Credits available	Response date	Relevant links	Status
Domain Error v2	Exploring Error and Model Domains in Materials Machine Learning	I am interested in exploring the accuracy of different error estimation methods and how they might be used to identify data as within or outside the domain of a model. The goal of this work is to develop automated approaches to obtaining reliable predictions, with uncertainty estimates, of engineering properties. We will use standard machine learning models in scikit-learn. We will focus on Bayesian approaches (e.g., Guassian Process Regression), ensembles approaches (e.g., bootstrap methods), outlier identification methods (e.g. Isolation Forests), and some domain specific approaches (e.g., using composition in materials systems). The results would be valuable to many engineers as more machine learned models are used to predict values. I expect this project to take 1-2 years and engage two teams of students of about 4 each, although the same students may not all stay the whole time. The project will provide opportunities for learning skills in machine learning and statistics, python programming, and teamwork, and we expect the work to yield one or more scientific publications and open source code packages. This is a "regression" type project.	Ongoing	10	None	Dane	Morgan	Materials Science and Engineering	ddmorgan@wisc.edu		Dane Morgan, MS&E, ddmorgan@wisc.edu	8	No	Yes	Ongoing		Active-Recruiting
Foundry Models	Cloud Based Prediction Tools for Materials Properties	I am interested in developing machine learning models for materials properties and sharing them as cloud resources through our recently developed Foundry infrastructure. The goal of this work is to make useful models available and showcase the infrastructure as a key resource for the materials community. We will use standard machine learning models typically from scikit-learn, tensorflow, and pytorch and typically apply gradient boosted trees or deep learning neural networks. The resulting models will provide values data for materials scientists and demonstrate a new way to share models and data for this community. An example might be developing integrated models and data for properties of small organic molecules. I expect this project to take 1-2 years and engage two teams of students of about 4 each, although the same students may not all stay the whole time. The project will provide opportunities for learning skills in machine learning, modest python programming, and teamwork, and we expect the work to yield one or more scientific publications and online models and databases.	Ongoing	10	None	Dane	Morgan	Materials Science and Engineering	ddmorgan@wisc.edu	608-234-2906	Dane Morgan, MS&E, ddmorgan@wisc.edu	8	No	Yes	Ongoing		Active-Recruiting
NLP Data Extraction	Materials Databases Generated by Natural Language Processing	I am interested in developing natural language processing (NLP) models to extract materials properties from published papers. The goal of this work is to generate useful databases for the materials community, as well as provide tools for others to generate such databases. We will use existing NLP frameworks, such as BERT or SciBERT with some retraining, to try and analyze the contents of scientific papers and extract sentences containing data. The resulting databases will provide values data for materials scientists and the approaches potentially transform how data is collected in the materials community. I expect this project to take 2-3 years and engage one team of students of about 4, although the same students may not all stay the whole time. The project will provide opportunities for learning skills in machine learning, python programming, and teamwork. I expect the work to yield one or more scientific publications and online models and databases. Meetings for this project are held online via MS Teams. This is a "regression" type project.	Ongoing	10	None	Maciej	Polakk	Materials Science and Engineering	mppolak@wisc.edu		Maciej Polak, MS&E, mppolak@wisc.edu	8	No	Yes	Ongoing		Active-Recruiting
Transformer for Chemistry	Flash Point BERT	BERT is the most popular NLP models used in the deep learning community now since it offers the unique ability to process sequential data. Flashpoint is an important material property and it has been studied by our group using a Graph-based neural network and a comprehensive SMILES string of compounds with corresponding flashpoint values has been collected too. Recently, people find that we can apply BERT to study the SMILES strings(https://arxiv.org/pdf/2007.16012.pdf), so an interesting question is that could we use BERT to investigate the SMILES string of flashpoint of compounds? This is a "deep learning" type project.	Ongoing	10	None	Ben	Afflerbach	Materials Science and Engineering	bafflerbach@wisc.edu	512-934-4497	Benjamin Afflerbach, MS&E, bafflerbach@wisc.edu	4	No	Yes	Ongoing	https://arxiv.org/pdf/2007.16012.pdf	Active-Recruiting
		A wealth of valuable research data is locked within the millions of research articles published every year. Reading and extracting pertinent information from those articles has become an unmanageable task for scientists. Moreover, these data are loosely structured, encoded in manuscripts of various formats, embedded in different content types, and are, in general, not machine accessible. Thus, studies that automatically leverage this valuable information are not tractable or even possible. Current approaches employ humans to manually extract data, define extraction rules, or annotate training corpora for machine learning approaches through tedious, time-consuming, error-prone and sometimes expensive processes. In the specific case of scientific information extraction, the need for pointed expertise increases costs and decreases the generalization of extraction methods. Our work so far has explored ways to identify sentences from artciles from which experts could identify the information to be extracted. We have two publications showing how we are able to drastically reduce the text to be reviewed by experts (6%) and achieve 100% recall. We are now looking deeper into identifying the information within these sentences using word vectors. This is a machine learning, natural language processing project, at the intersection of data science and materials science, which involves working python programming.	Ongoing	10	Python	Roselyne	Tchoua	Materials informatics	rtchoua@depaul.edu		Roselyne Tchoua, rtchoua@depaul.edu	2	No		Ongoing	https://via.library.depaul.edu/cdm_etd/25/
MAST-ML Coding	Development of the Materials Simulation Toolkit for Machine Learning	The Materials Simulation Toolkit for Machine Learning (MAST-ML) is an automated, open-source toolkit designed to acclerate data-driven materials research. We are interested in augmenting our current MAST-ML software with new analysis routines central to understanding the performance of machine learning models as they pertain to materials research. Some examples include a robust model error estimation framework, establishing domain of valid model performance, and materials-centric statistical tests, such as training and validating models on certain data subsets as defined, for example, by materials within some composition threshold. This project will provide opportunities for learning skills in machine learning, data analysis, python programming, and materials science.	Inactive	10	Python	Ryan	Jacobs	Materials Science and Engineering	rjacobs3@wisc.edu		Ryan Jacobs, MS&E, rjacobs3@wisc.edu	4	No	Yes	Ongoing	https://github.com/uw-cmg/MAST-ML	Inactive
Defect Detection (SOTA Models)	Defect Detection using State-of-the-art Transformer models	We are interested in using deep learning-based object detection models to characterize and quantify defects in electron microscopy images. Aquiring well-labeled training data of experimental images is very expensive, time-consuming, with a degree of subjectivity that may affect final model performance. Here, we want to train and assess state-of-the-art Transformer-base object detection models and compare performance against previous regional convolutional neural network (R-CNN) models. This project will provide opportunities for learning skills in deep learning, specifically object detection and image generation, data analysis, and python programming. This is a "deep learning" type project.	Inactive	10	Python	Ryan	Jacobs	Materials Science and Engineering	rjacobs3@wisc.edu		Ryan Jacobs, MS&E, rjacobs3@wisc.edu	2	No	Yes	Ongoing		Inactive
MRI Phantoms	Defect detection in MRI phantoms	Our team develops MRI test objects ( phantoms) consisting of vials with various gels in them. As part of an ongoing project we are developing new phantoms and optimizing manufacturing. One perpetual challenge with gel based phantoms is the presence of small air bubbles that create artifacts in MRI images. As quality control, we currently scan these vials (~50 at a time) in 3D using MRI and manually look for bubbles. We are interested in automating this process using a bubble-detecting algorithm (eg: DL based) to determine which vials are OK and which ones should be discarded. Upon solution of this problem, additional steps may be warranted, including automated measurement of the phantom MRI properties, etc.	Ongoing	10	Python	Diego	Hernando	Radiology and Medical Physics	dhernando@wisc.edu	608-265-7590	Diego Hernando, Radiology, dhernando@wisc.edu, Jean Brittain, Calimetrix, jbrittain@calimetrix.com	2	Possibly	Possibly	Ongoing	http://www.calimetrix.com http://qiml.radiology.wisc.edu	Inactive
Feature Functions	Exploring feature engineering techniques for MAGPIE elemental features	We are interested in improving existing techniques for feature generation and engineering in materials science machine learning applications. Previous research has shown that machine learning models can learn from elemental property information to predict materials properties. This method relies on models learning all of the underlying relationships between the elemental properties. Information might be learned more readily by combining or rearranging these elemental features to more directly represent the property being predicted. We are exploring various methods for manipulating these features using existing datasets that have been analyzed previously to see if we can improve upon previous work with these new feature engineering methods. This is a "regression" type project.	Ongoing	10	Basic Python / None	Ben	Afflerbach	Materials Science and Engineering	bafflerbach@wisc.edu	512-934-4497	Benjamin Afflerbach, MS&E, bafflerbach@wisc.edu	6	No	Yes	Ongoing		Active-Recruiting
BLADDER	A New Paradigm for Systems Physiology Modeling: Biomechanistic Learning Augmentation with Deep Differential Equation Representations (BLADDER)	A major unsolved challenge that cuts across many domains of computational science and engineering is to bridge the gap between data-driven machine-learning approaches and insight-rich mechanistic modeling by differential equations. ML scales with the data, but offers little or no mechanistic insight; in contrast, ODE models can incorporate physical, chemical, or biological mechanisms and yield insights into how systems work, but they don’t scale well with ‘big data.’ In the domain of human health, Stimulating Peripheral Activity to Relieve Conditions (SPARC) is a program of the National Institutes of Health (NIH), is seeking to bridge this gap. Our multi-investigator two-year project (Fall20 to Fall22) is bringing together computer scientists, mathematicians, biologists, and engineers to create a hybrid computational modeling approach, guided by data from laboratory experiments, linking ODE models using neural networks. Priority in the selection of student co-workers will be for those with experience or interest in ODE-modeling of physical, chemical, or biological systems and a strong foundation in computational science or engineering. More details: https://projectreporter.nih.gov/project_info_description.cfm?aid=10206953	ongoing	10	Python	John	Yin	Chemical and Biological Engineering, Wisconsin Institute for Discovery (WID)	john.yin@wisc.edu	608-316-4323	John Yin, CBE/WID, john.yin@wisc.edu	4	Possibly (summer 2021)	Possibly	Ongoing	https://www.youtube.com/watch?v=2zL2E50oRr8	Active-Recruiting
Edu	Introduction to Basics of Machine Learning Based Research	For students who wish to join the Skunkworks but have limited background in data science, machine learning, and related topics, we have an educational program. This program can be a great way for those with limited background to get up to speed. The program consists of about 10 modules that students work through during a semester with help from a mentor and their colleagues. The modules are very flexible and self-directed and the overall approach can be easily adjusted to specific needs. Students can finish the educational program early and move into a full research project if they wish.	9/20/21	10	None	Ben	Afflerbach	Materials Science and Engineering	bafflerbach@wisc.edu	512-934-4497	Benjamin Afflerbach, MS&E, bafflerbach@wisc.edu	0	No	Yes	9/27/21		Active-Recruiting
CNN-based Conductivity Prediction	Predicting the ionic conductivity of polymer-ceramic composite solid electrolytes using 3D Convolutional Neural Network	Currently used Li-ion Batteries usually consist of flammable liquid electrolytes. Nonflammable solid-state electrolytes have been proposed to replace the liquid electrolytes for making an all-solid-state battery that is safe, durable, and particularly attractive for powering large devices such as electric vehicles and drones. Although significant progress has been made through almost five decades of development, hallmarked by the discovery of various solid electrolytes with excellent Li-ion conductivity which allows for fast charging, it is still challenging to achieve a fast Li-ion transport across electrolyte-electrode interface. To overcome this barrier, a polymer-ceramic composite electrolyte has been proposed, which has the potential of enabling a fast Li-ion transport both in the electrolyte and across the electrode-electrolyte interface. The purpose of this project is to train a 3D convolutional neural network (CNN) model for achieving a fast and accurate prediction of the effective Li-ion conductivity of the composite polymer-ceramic solid electrolytes. It is a regression type task, and we will have the datasets of both the input “microstructure” of the composite and the target Li-ion conductivity “three numbers” ready. Extensive experiences in CNN regression type task are required. We already have a preliminary 3D CNN model for you to start with. Students with good writing skills are preferred. The student is expected to work in a relatively high degree of independence, complete the majority of CNN model training, testing, and data representation, and have a weekly discussion with the faculty advisor and a third-year PhD student. The student needs to be enrolled in independent study for credits. We will cover the expense needed to run CNN on Euler at WACC https://wacc.wisc.edu/.	ongoing	10	Relatively experienced in CNN regression; Good skills in writing science	Jiamian	Hu	Materials Science and Engineering	jhu238@wisc.edu		Jiamian Hu, MS&E, jhu238@wisc.edu	1	No	Yes	Ongoing		Pending
COoL	Chemical Origins of Life (COoL): a reaction network approach	The chemistry which led to the origin of life on the early Earth remains poorly understood. One of the major challenges is the combinatorial explosion that occurs when unspecific chemistries are used to synthesize sequence-specific polymers, like proteins. We aim to investigate the kinetics of amino acids forming peptides in prebiotic systems, a process containing hundreds of potential reactions and thus requires advanced network reconstruction tools. Similar tasks have been addressed by modelling tools; however, these tools have not yet been adapted to the specific constraints of origins of life systems. The goal of this project is to adapt the Rule Input Network Generator (RING) to study prebiotic peptide reactions. Students will be co-mentored by Prof. John Yin (UW Madison, Chemical and Biological Engineering) and Prof. Srinivas Rangarajan (Lehigh University, Chemical and Biomolecular Engineering). Priority will be given to students with an interest in computational network modeling and a background in chemical engineering or computational science.	ongoing	10	domain specific English-like reaction language	John	Yin	Chemical and Biological Engineering, Wisconsin Institute for Discovery (WID)	john.yin@wisc.edu		John Yin, CBE/WID, john.yin@wisc.edu	4	possibly summer 2021	yes	Ongoing	https://wid.wisc.edu/featured-science/origins-of-life-and-john-yin/	Active-Full
Atomic descriptors	Predicting defect energies in concentrated alloys	The formation energies of point defects in concentrated alloys vary upon the varying local atomic environment and usually exhibit a distribution. This raises a big challenge for determining the formation energies of point defects using atomistic calculations due to the finite simulation cell sizes. We are interested in training a machine learning model for predicting the thermodynamic and kinetic properties of point defects in concentrated alloys such as stainless steels. We will focus on the Gaussian Approximation Potential framework that uses the smooth overlap of atomic positions (SOAP) descriptor. The trained model based on atomistic calculation is expected to give efficient prediction of defect energies and help identify the factors that govern the distributions of defect energies.	ongoing	10	python; knowledge in crystal structure	Yongfeng	Zhang	Engineering Physics	yzhang2446@wisc.edu	608-890(3779)	Yongfeng Zhang, yzhang2446@wisc.edu	1	yes	No	ongoing	zhang.ep.wisc.edu	Active-Recruiting
Feature Selectors	Best practices for selecting and validating feature sets for accurate machine learning models	Building accurate and transferable machine learning models depends critically on the feature set used to numerically represent the data of interest. Here, we are interested in developing new practical methods of feature selection, with an emphasis on evaluating our methods on a number of experimental and computed materials property datasets. This work will leverage and build on the MAST-ML package, which was developed in our group to streamline and automate construction of machine learning models. This work will result not only in building practical machine learning and data science experience, but potentially result in a generally useful method for constructing robust feature sets for production of accurate models of numerous materials properties.	Ongoing	10	Python, machine learning	Ben	Afflerbach	Materials Science and Engineering	bafflerbach@wisc.edu		Benjamin Afflerbach, MS&E, bafflerbach@wisc.edu	4	No	Yes	Ongoing		Active-Recruiting
Automated atom finder	Deep learning to automatically find atom positions in electron microscopy images	Modern electron microscopes can readily generate images that see single atoms. This enables quantitative analysis of atomic structures to understand and design new materials, from better computer chips to new catalysts for environmentally friendly fuels. However, atoms are still generally found by human labeling of each atom in images. We seek to develop deep learning models to find atoms in images, integrating experimental and simulated training databases.Please see more details on the process of this project from this video of the talk in Microscopy&Microanalysis 2021.	Ongoing	10	Python, machine learning	Jingrui	Wei	Materials Science and Engineering	jwei74@wisc.edu		Paul Voyles, paul.voyles@wisc.edu	4	No	Yes	Ongoing		Active-Full
Liquid crystal analysis	Data-driven analysis of liquid crystal optical responses	Liquid crystal (LC) thin films are valuable materials for real-time sensing applications because molecular binding events cause LC films to exhibit rich, time-dependent changes in optical properties (optical responses) that provide information on the type and concentration of molecules that trigger the response (analytes). Optical responses lead to spatially varying colors and intensities that can be visualized as videos. The goal of this work is to build a data-driven computational toolkit for the analysis of LC optical responses. We will incorporate methods for image manipulation to facilitate data set curation, compare the ability of different machine learning methods (convolutional neural networks, recurrent neural networks, etc.) to relate optical responses to analyte properties, and implement methods to understand what features of the videos are most important for analyte identification. Our initial data set consists of hundreds of videos of optical responses triggered by the binding of environmental contaminants. We expect this project to involve collaboration between a team of undergraduate students for 1-2 years. Students will learn skills in machine learning, statistics, and Python programming, and will benefit from working in a collaborative environment involving five different principal investigators and their groups. The resulting toolkit will be valuable both for the analysis of environmental contaminants and more broadly to the analysis of LC-based sensor materials. We expect this work to be released as an open-source code package and a potential publication.	Ongoing	10	None; Python experience preferred	Reid	Van Lehn	Chemical and Biological Engineering	vanlehn@wisc.edu	608-263-9487	Reid Van Lehn, vanlehn@wisc.edu and Victor Zavala, zavalatejeda@wisc.edu	4	No	Yes	Ongoing		Active-Recruiting
ML tools for AM	Machine learning tool for identifying defect formation mechanisms from in-situ high-speed X-ray imaging data	Defect formation in metal additive manufacturing (AM, also known as 3D printing) severely hinders their applications because defects can cause failure of parts. Keyhole-induced porosity is one of the most common and detrimental defects in metal additive manufacturing. However, the formation mechanism is not fully understood. Here, the real-time keyhole porosity generation process during additive manufacturing has been experimentally recorded by high-speed high-resolution X-ray imaging. The first objective of this work is to develop an automated approach using machine vision to identify the keyhole (a depression cavity in liquid metal induced by intensive vaporization of materials during laser heating) and keyhole-induced pores out of thousands of X-ray images. The algorithm is expected to achieve fast and accurate marking/labelling of keyhole profile and pore profile in given X-ray images. After all the images were processed, the second objective is to use machine learning to explore the correlations between keyhole fluctuations and pore formation. This project will provide opportunities for learning skills in machine vision, machine learning, and X-ray imaging of additive manufacturing processes.	Ongoing	10	Python	Qilin	Guo	Mechanical Engineering	qguo46@wisc.edu	608-234-2906	Lianyi Chen, lianyi.chen@wisc.edu	8	Yes	Yes	Ongoing		Active-Recruiting
Glass Mechanical Properties	Machine Learning predictions of mechanical properties of metallic glasses	Metallic Glasses are a class of metals that can potentially have very interesting and useful properties for a range of applications. However, they have a large drawback in that discovery and tuning of new glassy compositions is time consuming and difficult experimentally. Therefore, we will train and assess models for predicting mechanical properties of glasses from simple compositional features, allowing experimentalists greater insight into the materials before sythesis. This can help guide and accelerate research by limiting searches to the most promsising materials based on our model's predictions.	Ongoing	10	machine learning	Benjamin	Afflerbach	Materials Science and Engineering	bafflerbach@wisc.edu	512-934-4497	Benjamin Afflerbach, MS&E, bafflerbach@wisc.edu	6	Yes	Yes	Ongoing		Active-Recruiting
Computational Design	Undergraduate Research in Compuational Design	Computational Design and Manufacturing Lab is interested in advancing computer methods for design and manufacturing applications. Current research activities in CDM Lab focus on machine learning and topology optimization for design and additive manufacturing applications. Example projects for undergraduates include the training of deep neural network for optimal design of heat sinks and heat exchangers, and the training of convolutional neural network for computer-aided design (CAD) modeling such as constructing NURBS surfaces.	Ongoing	10	Students with interest in mathematics and computer programing (Python and/or C++) are encouraged to apply.	Xiaoping	Qian	Mechanical Engineering	qian@engr.wisc.edu	608-890-1925	Xiaoping Qian, Mech E, qian@engr.wisc.edu		No	Yes	Ongoing		Active-Full
Kidney projects-Cancer	Deep Learning to identify aggressive tumor features in Renal cell carcinoma (kidney cancer)	As the volume of computed tomography (CT) performed for a variety of indications continues to increase, the incidence of renal cell carcinoma (RCC) has also continued to rise. Spatial heterogeneity is a common feature of RCC, with multiple studies demonstrating variability within tumors with respect to pathologic features, genomics, and RNA/protein expression. This heterogeneity gives rise to a spectrum of biologic and clinical behavior, with an increasingly less aggressive management approach in more indolent disease and a move towards nephron sparing approaches in cases where intervention is warranted. Pathologic markers of tumor aggressiveness such as higher nuclear grade or presence of sarcomatoid features may only be present in a small portion of the tumor but may profoundly impact treatment decisions and prognosis. These small areas can be challenging to identify on biopsy, and although radiomic features provide more global assessment and have shown some promise in capturing and characterizing tumor heterogeneity, some aggressive tumor features have remained elusive at imaging. The purpose of this project is to apply deep learning to previously segmented renal cell carcinomas on CT images with known pathologic features to help identify an imaging signature for aggressive features (high nuclear grade, sarcomatoid features). We have manually segmented approximately 124 large renal cell carcinomas (> 7cm) into volumetric regions of interest (ROIs) which represents the nuclear grade cohort. We have segmented a second data set of 45 RCCs with sarcomatoid features with 49 size matched non-sarcomatoid RCCs to serve as controls with extracted radiomics features on which we can perform similar analysis. The team would create a deep learning model to evaluate the CT images in this patient cohort with known kidney cancers. The model may require coding to incorporate the mask (image segmentations) with the CT image stacks showing the kidney tumor, or could also be run without the annotations on the date.	Ongoing	10	Machine learning, ?CNNs	Meghan	Lubner	Radiology, UWSMPH	mlubner@uwhealth.org	608-263-9028	Meghan Lubner (radiology), Andrew Wentland (radiology), E. Jason Abel (urology), Dane Morgan (Material Sciences)	0	Possibly	Possibly	Ongoing		Active-Recruiting
Kidney projects-benign masses	Machine and Deep Learning to predict clinical outcomes in angiomyolipomas	Angiomyolipomas are histologically benign tumors that occur in the kidneys, either sporadically or associated with syndromes such as tuberous sclerosis. They consist of fat, muscle, and abnormal vessels. Because of these abnormal vessels, these lesions can be at risk for spontaneous and at times, life threatening hemorrhage. Based on small clinical series, a size threshold of 4 cm has been associated with hemorrhage and need for treatment. We have collected a series of patients with angiomyolipomas imaged over time with at least one baseline computed tomography study (many pts in the cohorts have both imaging and clinical follow up). We have segmented the tumors and extracted a large number of radiomics features. This extracted data would be amenable to machine learning analysis to try to identify radiomic features that may predict bleeding risk or need for intervention. This could be performed with XGBoost models etc. In addition, this segmented image data could be used as input to a deep learning model that could be used to predict similar outcomes. In theory, a model similar to the project above could be re-trained and applied to this data.	Ongoing	10	MAchine learning, ?CNNs	Meghan	Lubner	Radiology, UWSMPH	mlubner@uwhealth.org	608-263-9028	Meghan Lubner (radiology), Andrew Wentland (radiology), E. Jason Abel (urology), Dane Morgan (Material Sciences)	0	Possibly	Possibly	Ongoing		Active-Recruiting
Short Name	Title	Project Description	Application deadline	Requirements: hrs/wk	Requirements:skills/courses/other	Contact: First Name	Contact: Last Name	Contact: Department	Contact: Email	Contact: Phone	Team leads (Names, Departments, Emails)	Number of Positions (max)	Funds available?	Credits available	Response date	Relevant links	Status