Variable-Dependent Predictability of Coastal Atmospheric Parameters Using Physics-Informed Machine Learning in the Marine Wave Boundary Layer
Online via Zoom
Dr. Mohammad Karim
mkarim@umassd.edu
https://umassd.zoom.us/j/98731680128?pwd=y211mggmO6iXQUy9DjXa5wbyPDcBqv.1
Faculty Supervisor:
Dr. Mohammad Karim, Electrical & Computer Engineering
Committee Members:
Dr. Donghui Yan, Mathematics
Dr. Tariq Manzur, Electrical & Computer Engineering
Abstract:
Advances in environmental sensing technologies have enabled the collection of high-frequency atmospheric observations, creating new opportunities for applying machine learning to short-term environmental prediction. Despite considerable progress in data-driven atmospheric modeling, comparatively little attention has been given to how atmospheric predictability varies among variables governed by different physical processes, how strongly prediction depends on atmospheric memory, and how boundary-layer regime modifies the predictive information available from instantaneous observations. Understanding these relationships is essential for developing forecasting systems that are both accurate and physically interpretable.
In this study, we systematically assess the short-term predictability of five key atmospheric variables—solar radiation, air temperature, wind speed, barometric pressure, and relative humidity—using approximately 2.2 years of high-frequency meteorological observations collected within the Marine Wave Boundary Layer (MWBL) at the School for Marine Science and Technology (SMAST), University of Massachusetts Dartmouth. To account for fundamentally different radiative forcing and boundary-layer dynamics, the observations were separated into daytime and nighttime regimes and analyzed independently. A unified physics-informed machine learning framework incorporating atmospheric memory, temporal encoding, and wind-vector decomposition was developed and evaluated using a strict chronological train-validation-test strategy to ensure realistic prediction conditions.
The results show clear differences in predictability and optimal model complexity across the five atmospheric variables. Air temperature and barometric pressure exhibited strong temporal persistence and achieved high predictive performance using linear models, whereas solar radiation and wind speed benefited from nonlinear ensemble-learning approaches capable of capturing more complex atmospheric behavior. Relative humidity achieved strong predictive performance but exhibited greater sensitivity to rapid moisture variability and evolving boundary-layer conditions. Additional experiments conducted comparing the complete framework with a reduced framework, in which temporal-memory features were removed, demonstrated that atmospheric memory contributes to the predictability of all five atmospheric variables, although its importance varies considerably among them. Comparisons between daytime and nighttime conditions further showed that boundary-layer regime modifies the predictive information available from instantaneous atmospheric observations while preserving the overall relationship between atmospheric physics and model complexity.
To evaluate practical applicability, model predictions were compared with independent observations collected from an ATMOS 41W all-in-one weather station deployed near the University of Massachusetts Dartmouth Campus Tower, providing the preliminary field-based validation of the proposed framework. The best-performing solar radiation prediction model was subsequently integrated into a prototype Streamlit-based real-time prediction dashboard, demonstrating a practical pathway from environmental data science research to operational coastal forecasting applications. Overall, this study demonstrates that atmospheric predictability is governed by the underlying physical processes of each atmospheric variable, that atmospheric memory provides the dominant source of predictive information, and that boundary-layer regime modifies predictive information without fundamentally altering the relationship between atmospheric physics and appropriate model complexity. These findings provide both scientific insight into coastal atmospheric predictability and a practical foundation for future environmental forecasting systems.
For further information, please contact Dr. Mohammad Karim at mkarim@umassd.edu.