Predictive scale-bridging simulations through active learning.

Area of Science:

• Computational Science
• Multiscale Modeling
• Active Learning

Background:

• Increasing computational power necessitates advanced methods beyond brute-force simulations.
• Accurate modeling of phenomena like transport in nanoporous media and inertial confinement fusion requires incorporating molecular-level interactions.

Purpose of the Study:

• To develop a novel active learning approach for optimizing fine-scale simulations to inform coarse-scale hydrodynamics.
• To enhance physical fidelity in computational models by bridging scales.

Main Methods:

• Utilizing active learning to guide the execution of local fine-scale molecular dynamics calculations.
• Forecasting continuum coarse-scale trajectories to speculatively initiate new fine-scale simulations.
• Dynamically updating coarse-scale models based on fine-scale results.
• Quantifying uncertainty within neural network models.

Main Results:

• A new capability for integrating multi-scale simulations has been developed.
• The approach effectively uses fine-scale data to enhance coarse-scale hydrodynamic predictions.
• Uncertainty quantification in the predictive models has been addressed.

Conclusions:

• The developed active learning framework offers a powerful strategy for improving the accuracy and efficiency of computational simulations.
• This method enables more precise predictions in fields requiring multiscale analysis, such as energy extraction and fusion research.