Combining machine learning with knowledge-based modeling for scalable forecasting and subgrid-scale closure of large,...

Area of Science:

• Dynamical Systems
• Machine Learning
• Computational Science

Background:

• Predicting large, spatiotemporally chaotic dynamical systems is crucial for applications like weather forecasting.
• Existing methods often struggle with large-scale systems and rely on imperfect models and historical data.

Purpose of the Study:

• To develop a scalable machine learning framework for predicting the time evolution of complex dynamical systems.
• To integrate historical time series data with imperfect models for improved predictive accuracy.
• To reduce the amount of training data required for effective model training.

Main Methods:

• A parallel machine learning prediction scheme was developed.
• A hybrid composite prediction system combining knowledge-based and machine learning components was employed.
• The approach was designed for scalability to very large and complex systems.

Main Results:

• The proposed method demonstrates excellent performance and scalability for large dynamical systems.
• Parallelization significantly reduces the time series data needed for training machine learning components.
• The scheme effectively incorporates subgrid-scale dynamics using training data for improved closure.

Conclusions:

• The combined parallel and hybrid machine learning approach offers a scalable and data-efficient solution for predicting chaotic dynamical systems.
• This method enhances the accuracy of predictions by effectively handling subgrid-scale processes.
• The findings have significant implications for fields requiring accurate forecasting of complex systems.