Accelerating multiscale modelling of fluids with on-the-fly Gaussian process regression.

Area of Science:

• Multiscale modeling
• Computational fluid dynamics
• Statistical learning

Background:

• Hybrid continuum-atomistic models are crucial for simulating fluid systems at multiple scales.
• Molecular dynamics (MD) simulations are computationally expensive, limiting their application in large-scale systems.
• Existing hybrid methods often require redundant atomistic simulations, leading to inefficiencies.

Purpose of the Study:

• To develop an accelerated hybrid model for multiscale fluidic systems.
• To leverage Gaussian process regression (GPR) as a surrogate for MD simulations.
• To enable on-the-fly uncertainty quantification and adaptive database augmentation.

Main Methods:

• Gaussian Process Regression (GPR) was employed as a surrogate model for MD simulations.
• The GPR model predicts atomic-scale information from continuum model inputs and outputs.
• An adaptive strategy was implemented to perform new MD simulations when prediction uncertainty exceeds a threshold.

Main Results:

• The hybrid scheme accurately predicts atomic-scale information, achieving accuracy near thermal noise levels for low uncertainty thresholds.
• Significant computational speed-ups (up to an order of magnitude) were achieved compared to full MD simulations.
• The trade-off between accuracy and computational efficiency can be tuned by adjusting the uncertainty threshold.

Conclusions:

• The proposed GPR-based hybrid scheme offers a substantial improvement in efficiency for multiscale fluidic simulations.
• This method reduces the reliance on numerous, often repetitive, MD simulations.
• The tunable accuracy-efficiency balance makes the scheme adaptable to various research needs in fluid dynamics.