Quantifying Uncertainty in SOA Yield Predictions and Associated Health Risks using AutoML and Quantile Regression

Poster of work completed in Google DeepMind Research Ready Internship at University of Manchester. Abstract: Predicting the mass yield of secondary organic aerosol (SOA) is a critical challenge in atmospheric science, creating major uncertainty in air quality and climate models. In this study, we use Automated Machine Learning (AutoML) to build an accurate estimator for SOA mass yield based on complex chemical simulation data. We then extend this model using quantile regression to quantify predictive uncertainty, generating a 90% prediction interval rather than single-point estimates. By translating these probabilistic yield predictions into a Mean Attributable Mortality (AMSOA) metric, we demonstrate a novel approach for conducting more robust, uncertainty-aware health impact assessments.