Aerosol Thermodynamics Neural Network System (ATNNS): A Python Implementation of A Mixture of Experts Neural Network Framework for Predicting Inorganic Aerosol Thermodynamics

This work proposes a Mixture of Experts (MoE) neural network framework to predict E-AIM Model IV outputs, including water content and vapor pressure products for ammonium nitrate and chloride. The system, which utilizes a decision-tree-based routing strategy, achieves an average Mean Absolute Percentage Error (MAPE) of 1.7% across all models. These results highlight how physics-aware partitioning and data transformations can enable accurate, scalable approximations of thermodynamic systems.