Machine Learning-Based Estimation of Superdroplet Growth Rates
Published:
Project Overview
Droplet growth and size spectra play a crucial role in the microphysics of atmospheric clouds. However, it is challenging to represent droplet growth rate accurately in cloud-resolving models such as Large Eddy Simulations (LESs). To overcome the limitations of the “well-mixed” assumption in traditional LES solvers, this project proposes a parameterization for superdroplet growth using high-fidelity Direct Numerical Simulation (DNS) data. A novel clustering algorithm maps droplets in DNS fields to superdroplets, and a machine learning model is developed to relate the effective growth rate of these superdroplets to filtered DNS flow variables.
Key Features
- Advanced Data Processing: Ingests and transforms raw Lagrangian and Eulerian data from DNS simulations into a structured format for machine learning.
- Feature Engineering: Implements multiple scaling methods to extract and prepare features and labels for model training.
- Machine Learning Model: Trains and validates a model to predict the effective supersaturation for superdroplets, achieving a promising R² value of nearly 0.9.
- A Posteriori Analysis: Includes scripts for evaluating the model’s performance and generalization on unseen data.
- Modular Workflow: The project is organized into distinct stages: post-processing, training, and a posteriori analysis.
Technical Details
- Language: Python
- Core Modules:
post-processing: Scripts for data ingestion and transformation.training: Scripts for feature extraction, model training, and validation.aposteriori: Scripts for a posteriori analysis and performance evaluation.
Publication
The findings and methodologies of this project are detailed in a preprint available on arXiv: https://arxiv.org/abs/2410.13890