Crop yield estimation
During my internship at SI Analytics in South Korea, I worked on a crop yield prediction project with the objective of predicting soybean and corn yields in the midwestern US based on satellite images and climate data. My role was to implement techniques for uncertainty quantification based on neural network approaches. I developed a method of uncertainty quantification using deep ensembles of LSTM networks that was also able to decompose the uncertainty into its aleatoric and epistemic components. This type of uncertainty analysis is very useful for gaining an intuitive understanding of the model and data in terms of the prediction errors.
While these models use full-year datasets to predict crop yields for the same year, data for the full year will not be available in practice. I also worked on developing time series models to extrapolate input features, histogram means, and climate data to the end of the year so that they can be fed into the existing neural network framework.
Webfoil: airfoil database, analysis, and optimization
During the summer months of 2018, I did research at the University of Michigan MDO lab. We developed a web application called Webfoil that serves as a comprehensive, open-source database of airfoils that includes built-in tools for aerodynamic analysis and optimization. The database includes downloadable information such as geometric coordinates and other airfoil characteristics. One of my contributions was to run various airfoil shapes through CFD software such as Xfoil and ADFlow and compile the lift, drag, and moment data into the database. The analysis tool contains several plots that change based on user input of Reynolds number, Mach number, and angle of attack. The optimization tool uses a surrogate model to generate a new airfoil shape that minimizes drag for a given thickness, Mach number, and coefficient of lift.
Autonomous can sorting machine
For an undergraduate engineering design course (AER201), my team designed and constructed a fully autonomous machine that is capable of sorting cans for recycling. I was the electrical member, responsible for the implementation of digital and analog interfacing electronics, including circuit design and sensor selection and implementation, and integrating with my electromechanical and microcontroller team members. Our machine placed 1st in the full semster course competition based on both performance and design.
There were four categories of cans: tin cans with label, tin cans without label, aluminum cans with tab, and aluminum cans without tab. Our machine used a dual level approach to first sort based on size. The upper level contained cutouts to allow the aluminum cans to fall to the lower level. A camshaft was used to shake both levels to separate the tin and aluminum cans, which then travelled into their respective chutes. At the bottom of the chute for tin cans, the presence or absence of a paper label on each can was evaluated using conductive thread. At the bottom of chute for aluminum cans, the presence or absence of a tab was detected using flex sensors. Based on the sensor data, the cans were dropped into the appropriate bucket.