Math-to-Industry Boot Camp IV

• Mentor Zeinab Takbiri, Cargill, Inc.
• Sijing Liu, Louisiana State University
• Damilola Olabode, Washington State University
• Puttipong Pongtanapaisan, The University of Iowa
• Nathan Willis, The University of Utah

Cargill is a major grain trader in the US. We utilize over 100,000 rail cars per year to ship grains to our domestic and export customers. Cargill uses railroad-supplied cars to move a lot of these shipments of grain. The railroads require us to take on an obligation to run their cars for a year. We are looking for help in developing a supply and demand model that can determine how many cars Cargill should take on in a given year as well as a forecast of the overall market’s need for railroad owned equipment.

• Mentor Jesse Berwald, D-Wave Systems
• Sarah Chehade, University of Houston
• Davood Damircheli, Mississippi State University
• Kevin Marshall, University of Kansas
• Li Zhu, University of Nevada

A quantum processing unit (QPU) is a programmable chip that leverages superposition and entanglement, fundamental quantum mechanical properties, to solve problems. The D-Wave quantum annealing computer currently operates with a 2048-qubit QPU. Calibrating such a chip in the presence of thermal, quantum mechanical, and design-specific noise is a critical component to producing a working quantum computer. 

D-Wave Systems has developed many internal calibration tests to infer anomalies observed in the QPU. Error correction on many levels is used to mitigate these anomalies wherever possible (though thermal and quantum fluctuations will always be present). The variety of tests often requires different models and statistical methods. This project looks at a test of a specific configuration of randomly coupled qubits (C4 RAN1). Students will implement and fit a model based on observations from the QPU. A significant part of the pipeline will include a visualization component to enable easy, and deeper, analysis of anomalies if they are present. 

• Mentor Nicole Bridgland, World Wide Technology
• Mentor Jonathan Hahn, World Wide Technology
• Steven Dabelow, University of Notre Dame
• Jordyn Harriger, Indiana University
• SeongHee Jeong, Louisiana State University
• Kiwon Lee, The Ohio State University

Mines have lots of moving parts, and timing of delivery between them is crucial.  Time that mining equipment spends idle represents lost production opportunity. Time trucks spend idle, while not as obviously problematic, represents at least wasted fuel if not lost production opportunity elsewhere in the mine.  Given a system of several shovels and crushers, and trucks moving material between them, how can you best decide where to send empty/loaded trucks as they become available? When equipment experiences delays, when should you reroute trucks vs simply wait it out, and how should you reroute them? The goal of this project will be to develop tools to help human dispatchers make these decisions, possibly in the form of machine-generated recommendations.

• Mentor Yiqing Cai, Gro Intelligence
• Xing Ling, Michigan Technological University
• Ben Strasser, University of Minnesota, Twin Cities
• Rahim Taghikhani, Arizona State University
• Tianyu Tao, University of Minnesota, Twin Cities

Gro is a data platform with comprehensive data sources related to food and agriculture. With data from Gro, stakeholders can make quicker and better decisions, which in most cases are time sensitive. In this project, the students will use data from Gro to identify analogous events. For example, people can compare and find a year with similar precipitation and soil moisture patterns to draw inferences about second and third order effects such as flooding or decreased crop planted area. This type of analysis can help quantify the impact of an event, and remedy the negative impact if it is severe and not avoidable.

Data will be provided through Gro API. Data pulled from Gro are in the format of time series, which are called data series. Different data series can come from different sources, and have different frequencies. For example, there is daily Precipitation data from TRMM, and NDVI at a frequency of 8 days (a type of vegetation index) from GIMMS MODIS.  

Goals: The deliverables of this project will be in the form of an executable model. Given a data series (or a set of data series), and a selected time period, find analogous periods in history that are most similar to this selected period. Given the project goal, it all boils down to defining similarity between a pair of data series, or concatenated data series. 

• Mentor Yanhua Yuan, ExxonMobil
• Dilek Erkmen, Michigan Technological University
• Anthony Nguyen, University of California, Davis
• Samantha Pinella, University of Michigan
• Jose Pedro Rodriguez Ayllon, University of Houston
• Nandita Sahajpal, University of Kentucky

Acquisition of seismic data in marine environment is a costly process. Traditionally, in marine seismic surveys, a boat tows a line of receivers while moving slowly. To obtain signals at the receivers, a wave source, typically an air gun, is generating a pulse with frequencies in the 10 of Hz which penetrates the earth and reflects back on the different layers of the earth. Recently, an innovation in this space was introduced that has been shown to have substantial savings and allowed for wider distances between the source and the receivers. In the new method, more than one seismic sources or air guns are fired with short or zero delays between them so that the signal generated by each source overlap at some or all receivers. The collected signals at the receivers are therefore blended together in simultaneous-source acquisition, and a “deblending” process is usually needed to separate signals from the individual sources before any further analysis. To make it easier for decoding, multiple sources are usually fired at a random time, and (or) with signatures coded differently. Based on the incoherence assumption, the deblending problem can be explored in different ways, including as signal processing problem, inversion problem, or data analytics problem. In this project, we will try these methods and look for a robust deblending algorithm to reconstruct individual source signals from encoded data.

• Mentor Jonathan Hill, ITM TwentyFirst LLC
• Brendan Cook, University of Minnesota, Twin Cities
• Vikenty Mikheev, Kansas State University
• Bibekananda Mishra, University of Kansas
• Negar Orangi-Fard, Kansas State University
• Matthew (Jake) Roberts, Michigan Technological University

Medical underwriting is expensive and time-consuming, involving trained underwriters who manually review medical history and long delays waiting for documentation. For these reasons, researchers in life insurance and related industries are fervently searching for methods to estimate mortality risk faster and at lower cost.

One proposed solution is to use a smaller set of medical features than what is typically collected in underwriting. These features could be collected through a questionnaire and used to generate a rapid estimate of mortality risk. This solution could have additional value in cases of full underwriting where some medical data is missing. A key objective will be quantifying the increase in uncertainty, or decrease in precision, as a consequence of using a smaller feature set.

During this week-long project, you will take a crash course in survival analysis, explore models for time-to-event data (including traditional and machine learning approaches), determine appropriate metrics, engineer features, and compete to create the best possible model of mortality risk. If time allows, there may be opportunity to develop novel modelling techniques.

We will be using a unique world-class dataset on senior life outcomes provided by ITM TwentyFirst, a Minneapolis-based life settlements servicing company.