Operator learning for predicting multiscale bubble growth dynamics.

Area of Science:

• Computational physics
• Machine learning
• Multiscale modeling

Background:

• Simulating multiscale problems across spatiotemporal scales is challenging.
• Current methods often rely on complex interface algorithms.
• Deep neural networks (DNNs) have not been systematically applied to these problems.

Purpose of the Study:

• Develop a deep learning framework for multiscale modeling.
• Investigate the capability of DeepONets to learn dynamics across different scales.
• Simplify multiscale modeling by avoiding fragile interface algorithms.

Main Methods:

• Utilized operator regression with DeepONets.
• Generated macroscale data using the Rayleigh-Plesset equation with Gaussian random fields (GRFs).
• Generated microscale data using dissipative particle dynamics (DPD) for stochastic processes.

Main Results:

• DeepONets accurately predicted macroscale bubble growth dynamics, outperforming LSTMs.
• DeepONets demonstrated accurate extrapolation outside the input distribution with minimal data.
• Trained DeepONets accurately predicted mean bubble dynamics from noisy DPD data.

Conclusions:

• DeepONets can unify macroscale and microscale models for multirate bubble growth.
• Operator regression via DNNs offers a novel approach to multiscale problems.
• DeepONets simplify modeling with heterogeneous descriptions in complex systems.