Frances Uy was born and raised in Honolulu, Hawai‘i and graduated from Moanalua High School. She is currently pursuing a B.S. in Computer Science on the Data Science Track at the University of Hawai‘i at Manoa. Frances values the collaborative and curious nature of Data Science, Machine Learning, and Artificial Intelligence. Alongside her studies, Frances fuels her passion in Computer Science as an ICSpark mentor on intro to web development for 6-12th graders. In her free time, Frances enjoys reading, pickleball, and spending time with family and friends. Her goals are to pursue postgraduate education and a career in Data Science. She looks forward to making a positive impact in our community through data-driven solutions.
Home Island: O’ahu
High School: Moanalua High School
Institution when accepted: University of Hawai’i at Manoa
Project Site: Gemini Observatory, Hilo, Big Island HI
Mentor: Hawi Stecher
Project Title: Classifying Cloud-Based Threats to Telescope Operations with Machine Learning
Abstract:
The Gemini North Telescope, located at the summit of Maunakea, enables many scientific discoveries due to the stable atmosphere above the 14,000-foot summit area and its close proximity to the equator. However, clouds can form above the summit level at any given time of the year and completely cover the telescope. Closing the telescope domes when bad weather-inducing clouds are present is one of the key tasks that Gemini telescope operators are charged with. Currently, operators at the Gemini Hilo Base Facility control room solely depend on still images from cloud cameras on the observatory roof to gauge cloud formation threats. This project aims to enhance the operators’ ability to identify cloud-induced weather risks in advance by leveraging the computational efficiency of a pre-trained ResNet50 model. This transfer-learned machine learning model is capable of classifying clouds into four categories that operators can use to determine when to close the telescope. The model will be trained to generalize various patterns within a real, manually labeled dataset of 200 images to maximize its ability to classify new cloud camera images accurately, and will be implemented into a real-time inference system in conjunction with the cloud camera server. For every cloud camera still taken 20 seconds apart, operators will be equipped with its respective classification category to complement their manual observation. The improved weather monitoring and predictive system will minimize the impacts of downtime caused by unexpected cloud cover, and upon successful testing, validation, and deployment, the system will be ready for immediate implementation by Gemini staff.