CUR - The Council on Undergraduate Research

Blue Mars Initiative: Developing Linear Regression and Artificial Neural Network Models to Forecast Mesoscale Martian Weather Conditions

Jared Frazier and Dr. David Butler, Department of Computer Science, 1301 E Main St, Murfreesboro, TN 37132

At any given moment, a devastating cosmic event could wipe all life on Earth from existence. In combination with pressures humanity places on Earth’s biosphere, extinction may be inevitable.¹ Going beyond our domain, further from the sun, and to the terrestrial planet Mars may be one way to reduce the possibility of human extinction.² Despite this lofty goal, the hostile Martian weather conditions differ vastly from those on Earth, and the ability to predict those conditions would be invaluable for successful colonization. In particular, the extremely wide range of temperatures (20°C to -73°C) are a significant barrier to implementing human infrastructure.¹ Traditional weather prediction techniques implemented on Earth such as numerical weather prediction (NWP) are extremely computationally intensive and are not always stable due to the volatile physical conditions of the Earth’s atmosphere. Additionally, NWP can not be easily transferred to predicting Martian weather.^2,3 To overcome this barrier, supervised machine learning—a method that is resistant to the incomplete understanding of atmospheric conditions that introduces uncertainties to NWP—is ideal for the even less understood Martian atmosphere.⁴ Weather data for Mars’ Gale Crater was collected by NASA’s Curiosity Rover and is available through their Planetary Data System. Two types of machine learning algorithms will be implemented for the prediction of mean temperature using Curiosity’s data: linear regression and artificial neural networks. These machine learning paradigms were selected due to the ability of each to account for the mix of non-linear and linear responses in weather.^5-7 For both models, ~3 Martian years of weather data will be used to predict ~1 year of test data. The mean and median absolute error for the prediction of mean temperature will be calculated and the models will be compared.   

Additional Abstract Information