Air Pollution Modeling with Probabilistic AI

During my studies at ETH Zurich, I developed a probabilistic machine learning model aimed at predicting urban air pollution levels using Gaussian processes.

Unlike standard AI approaches, probabilistic AI explicitly models uncertainty, making predictions more reliable by providing not only an estimated value but also the associated confidence level.

I implemented a regression framework based on kernel-based regression to capture spatial correlations between sensor locations and to interpolate missing data across the city.

This was followed by a fine-tuning phase during which I wrote programs for the search of optimal hyperparameters.

The goal of this project was to explore how to make AI models inherently more reliable by integrating uncertainty quantification at the core of their design.

Python, GPy, GPyTorch, NumPy, SciPy, Matplotlib, Scikit-learn, Git, GitHub