The following are projects completed by students in EEPS-DATA 1720: Tackling Climate Change with Machine Learning at Brown University in Spring 2024.
The projects below are shared with permission of the student authors and should not be redistributed.
Research Paper: Forecasting Air Quality Metrics with Evolved Recurrent Neural Networks by Aidan LaBella
The main question my project attempts to answer is "Can we forecast air quality data at timestep t + 1, even when sensor data is unreliable and/or not present?" This project, at a high level, is to train a model that is effective at predicting air pollution given certain parameters. I also propose using a toolkit for neural architecture search while training the model. The applications of a model that can predict air quality vary from being able to estimate the pollutants in the air based on other parameters, to being able to accurately forecast when the pollution will reach its high/low levels, much like a weather forecast.
[Final Paper] [GitHub] [MERRA-2]
Tutorial: Road-based Providence Pollution Prediction by Ayushman Choudhury and Bradley Marx
The question that we address is whether we can estimate the AADT (Average Annual Daily Traffic) counts of roads within a road network, based on publicly available information like satellite imagery of the road network and data on the types of roads within the network. We will be addressing this by re-implementing an existing paper [1] as a Google Colab tutorial. Measuring AADT is important, because it quantifies the volume of transportation, which is a significant part of a city's CO2 emissions, which are a major driver of global climate change. Quantifying these emissions are important for cities that wish to reduce them, so that they can gain a better understanding of the issue, diagnose specific problems in their city, and track progress towards decreasing transportation emissions. While many cities/regions have defined goals for their greenhouse gas (GHG) emission reductions, the ability to quantify their progress across all domains remains a challenge. Our tutorial demonstrates how to train a deep learning model to predict Log-AADT as a proxy for greenhouse gas emissions.
Click here to open tutorial (Google Colab)
[1] D. Rollend et al. “Machine Learning for Activity-Based Road Transportation Emissions Estimation,” in NeurIPS 2022 Workshop on Tackling Climate Change with Machine Learning, 2022.
Review & Tutorial: Navigating the Carbon Implications of Machine Learning by Julian Dai, Michael Fu and Tabitha Lynn
In recent years, the use of ML has grown exponentially, and its energy-intensive workload has increased with it. In fact, training one ML model can emit more than 626,000 pounds of carbon dioxide—nearly equivalent to the lifetime emissions of 5 cars. As ML becomes even more prevalent, the carbon impact of training ML models threatens to have a tangible impact on the global climate crisis. Tracking and disclosing the carbon impact of ML is a necessary step to raise awareness and ultimately mitigate its effects. In this work, we assess the current state of carbon trackers and predictors for ML models. We then detail ways to mitigate the carbon impact of models through improving efficiency, before proposing recommendations to incentivize the adoption of ML trackers.
Review: A Review of Prototype-based Explainable AI Studies for Potential Applications in Geosciences by Anushka Narayanan
Building inherently interpretable models can help alleviate the concerns with the commonly used post-hoc XAI techniques in the geo-sciences that have shown to produce variable results dependent on the particular technique. Prototype-based XAI methods are intrinsic methods for XAI that relies on comparing input data with a set of learned prototypes that are representative of the training data and computing a similarity metric to produce a prediction. In this work, we discuss a series of case-studies in the emerging field of prototype-based XAI methods that show potential for geo-scientific applications from three categories; the development and visualization of prototypes, types of prototypes and using prototypes for various learning tasks. For the case-studies we discuss how the authors use prototypes, their novel contribution and any limitations or challenges that may arise when adapting these methods for geo-specific tasks. We highlight various geo-scientific applications that may benefit from using or modifying these prototype-based XAI techniques.
Proposal: Carbon Capture Development using MOFs and GNNs by Anna Lapre
Excess greenhouse gasses are one of the largest contributors to global warming, which already has had a devastating effect on human beings, biodiversity, and the Earth more broadly. The amount of carbon dioxide specifically, which has been emitted in incredibly large quantities since the Industrial Revolution, has served as a catalyst for the Earth’s warming; yet, the emissions seem far from decreasing quickly enough to prevent temperatures from crossing irreversible thresholds of damage. A potential solution for this problem is Carbon Capture and Storage (CCS); carbon capture is a method for removing or reducing the amount of greenhouse gasses in the atmosphere via some chemical process. One such process is to use the cage-like structure of Metal Organic Frameworks (MOFs) to capture carbon by trapping carbon dioxide molecules while letting other atmospheric gasses simply pass through the MOF. The challenge is to find the MOF(s) which are best able to capture carbon, whether that be by having the largest capacity for the amount of CO2 molecules it can hold or how selective the structure is in trapping CO2 compared to other molecules. This, then, leads to the application of machine learning, particularly methods which easily depict molecules such as Graph Neural Networks (GNNs), to find the most optimal method for capturing carbon. This proposal outlines a project in which deep learning algorithms, particularly GNNs and generative GNNs, are used to search for the most effective MOFs, compared with the existing literature, in order to best capture carbon.
Proposal: Identifying Extreme Precipitation Days in Rondonia using a CNN by Caleb Ukaonu
The Amazon basin receives some of the highest amounts of precipitation in the world, but this precipitation varies spatially and temporally over the region. Observations also show that the spatial patterns of rainfall have changed over the past few decades (Paca et al, 2020), most likely due to deforestation and other land use change (Rizzo et al, 2020). [...] In order to gain insight on the capability of machine learning to predict extremes in this region, I will train a Convolutional Neural Network (CNN) on ECMWF Reanalysis 5 (ERA 5) data and CHIRPS satellite precipitation data to classify extreme precipitation days and non-extreme precipitation days. Understanding how well it can identify the extreme days will show how well it can identify future extreme precipitation days. The model will be trained on daily averaged data from November to March (i.e. the rainy season) 1981-2012, validated on data from 2012-2015 and tested on data from 2016-2019. Though there haven’t been any studies like this in Rondonia (southwestern Brazil), there have been many machine learning based prediction experiments that this study will be based on.