The following are projects completed by students in EEPS-DATA 1720: Tackling Climate Change with Machine Learning at Brown University in Spring 2026.
The projects below are shared with permission of the student authors and should not be redistributed.
Research Paper: Physics-Informed Hurricane Track Prediction by Brian Kao and Qizhi Sun
Accurate hurricane track prediction is critical for reducing societal impacts, but traditional Numerical Weather Prediction (NWP) is computationally intensive and contains parameterization errors. Machine Learning prediction is a powerful complement to NWP, though it may result in tracks that contradict physical laws. In order to improve the accuracy and realism of the predictions, the study introduces a physics-informed neural network (PINN) that incorporates the speed into the loss function. While PINN does not necessarily improve the distance accuracy compared with the Random Forest baseline, the shape of predicted tracks were substantially more consistent with the actual observations of hurricanes with significantly lower speed error. The results demonstrate a trader off between accurate location prediction and physical speed consistency.
Research Paper: Predicting and Categorizing Particulates in Providence: Using ML Methods to Understand PM 2.5 Regimes and Predict PM Levels with Wind Data by Jack Lardner
PM 2.5 is a dangerous air pollutant that has a variety of sources, and is of public health concern in Providence, Rhode Island. This project aims to understand the relationship between wind conditions and the concentration of PM 2.5 throughout the city using several pollutant and wind sensors that are part of the Breathe Providence air sensor network. To achieve this, both unsupervised and supervised machine learning methods were used. A clustering technique (K-means) was used to inform on existing wind-pollutant regimes, with low to moderate success. The clusters were mainly based on wind speed, and high concentration clusters did not see unique wind directions. Thus the supervised learning method (a simple neural network) could not perform well on PM 2.5 bin prediction using only wind speed. When including other factors like PM 2.5 lags, relative humidity, and time of day, the model was able to perform better (around 75% accuracy), with permutation importance indicating wind had relatively high predictive power compared to other factors. However, permutation performance and comparison to a baseline using only the one-hour lag feature indicates that the model’s performance was solely based on this one-hour lag and revealing that PM 2.5 concentration is autocorrelated than directly dependent on meteorological conditions. This project serves as a proof-of-concept, and can be expanded in several ways, including using stronger machine learning methods to predict the PM 2.5 concentration directly.
[Final Paper] [GitHub]
Research Paper: Tail-Weighted Residual Diffusion for Downscaling of Heat Extremes by Thomas Lin
Heat extremes are among the most damaging weather hazards, yet data-driven weather and downscaling models often underrepresent the most intense events. This paper investigates whether a tail-weighted loss function within a two-stage conditional residual diffusion framework can improve the downscaling of heat extremes. The proposed model first uses a deterministic U-Net to predict high-resolution 2-m temperature from ERA5 conditioning fields, then applies a conditional denoising diffusion model to predict the residual between the U-Net output and HRRR target temperatures. A tail-weighted diffusion loss upweights pixels exceeding a per-pixel 95th percentile temperature threshold, with sigmoid smoothing and a timestep-dependent noise taper. The study focuses on hourly May–September temperature fields from 2015–2022 over a subset of the Southern Great Plains. Results indicate that the tail-weighted model improves performance on extreme heat events, but performs worse on the general data distribution. These findings suggest that tail-weighting can improve the representation of extremes in diffusion models, but it is insufficient for reliable heat extreme downscaling.
Review Paper: Investigating the Current Climate and Weather Emulator Evaluation Paradigm by Aalyaan Ali
As climate science shifts towards democratization, Climate Emulators (CEs) have emerged as essential, low-cost alternatives to computationally intensive numerical solvers. However, the utility of these models is fundamentally limited by the reliability in their ability to generalize to new conditions and dynamics. This is especially important in a warming climate where this will only become more common. This work categorizes evaluation efforts into four distinct tiers, and discusses different examples of the latter three. This synthesis illustrates that much is still required to be done to ensure a standardized and rigorous evaluation of climate emulators to ensure reliability and validity. Future work should aim for more high-tiered evaluation.
Review Paper: A Review of Climate Model Evaluation Criteria in the Context of Machine Learning by Jacob Hirschhorn and Kyle Wisialowski
Since the release of the most recent report from the International Panel on Climate Change (IPCC), there has been a large increase in the number of publications that use machine learning (ML) to develop climate models, from roughly 35 per year in 2015 to 600 per year in 2020. It is estimated that this number has continued to increase given recent advances in machine learning algorithms and computational resources. These machine learning techniques have helped advance the field of climate modeling significantly. Due to their lower computational demand, they have allowed for more efficient ensembles of models to give more accurate measurements of uncertainty in predictions and have enabled better downscaling of climate models to show results from a regional scale rather than a global one, among other uses. However, they also have their own concerns, separate from established numerical models, such as their ability to generalize outside of their training distributions, their physical consistency, and their interpretability to human researchers. Given the broad application of these models, it is important to have a strong understanding of the methods used to evaluate their ability to ensure these models are as accurate and helpful as possible. [...] This paper examines the existing evaluation metrics from the IPCC that were detailed in 2013, the most recent examination of these techniques, which are directed towards numerical climate models, and makes recommendations for what other methods should be implemented.