Priscila Madrid Arroyos
MIT Department: Earth, Atmospheric, and Planetary Sciences
Faculty Mentor: Prof. Paul O’Gorman
Research Supervisor: Griffin Mooers
Undergraduate Institution: University of Texas at El Paso
Hometown: Santo Tomas, Guerrero, Chihuahua, Mexico
Website: LinkedIn
Biography
Priscila Madrid Arroyos is a rising senior and international student from Mexico,
pursuing a bachelor’s degree in Computer Science with a concentration in Data Analytics at the University of Texas at El Paso (UTEP). During her sophomore year, she interned as a
software quality assurance engineer at Eaton’s Thomas A. Edison Technical Center. Currently, she conducts research at the W.M. Keck Center for 3D Innovation. In the upcoming
academic year, she is excited to optimize the website for the UTEP Society of Hispanic
Professional Engineers (SHPE) chapter as the webmaster chair. At MSRP, she is working on evaluating neural network architectures’ effectiveness in accurately representing convective
dynamics within climate models. Priscila plans to pursue a Ph.D. in Computational Earth, Atmospheric, and Planetary Sciences to enhance crop productivity and agricultural resilience using machine learning and data-driven approaches. In her free time, she enjoys baking,
running, and playing basketball.
Abstract
Engineering Convolutional Neural Networks to Enhance Climate Models
Priscila Madrid1, Griffin Mooers2, and Paul O’Gorman2
1Department of Computer Science, University of Texas at El Paso
2Department of Earth, Atmospheric, and Planetary Science, Massachusetts Institute
of Technology
Accurate climate modeling is crucial for predicting crop yields and ensuring global food security. Traditional models, like General Circulation Models (GCMs) and Earth System Models (ESMs), have advanced the field but face limitations in computational efficiency and representing small-scale convection. Recent machine learning (ML) techniques, including random forests and neural networks (NNs), show promise in enhancing these models. However, Convolutional Neural Networks (CNNs) are particularly well-suited for models involving structured grid-like data, resulting in better performance and faster training times for many applications in computer vision. My research aims to engineer a CNN, replacing the current fully connected networks in predicting subgrid-scale moist convective tendencies over land and ocean surfaces. My main finding is that stochastic pooling can result in more robust feature extraction, which is vital for real-time prediction accuracy and long-term climate forecasts. The current status involves training the CNN using extensive climate model simulations and iterative hyperparameter tuning to refine performance. Improved weather predictions, such as convection, contribute to increased food security and reduced agricultural costs.
