Micro-macro consistency in multiscale modeling: Score-based model assisted sampling of fast/slow dynamical systems

Area of Science:

• Computational chemistry
• Materials science
• Multiscale dynamical systems

Background:

• Efficiently sampling phase space in multiscale dynamical systems is challenging.
• Systems can become trapped in local free energy minima, hindering direct simulation.
• Physics-based methods exist for enhanced sampling beyond free energy barriers.

Purpose of the Study:

• To investigate the use of score-based generative models (SGMs) for enhancing sampling in multiscale dynamical systems.
• To explore the coupling of machine learning generative models with physics-based enhanced sampling techniques.

Main Methods:

• Utilizing score-based generative models (SGMs) as a machine learning approach.
• Developing a framework for coupling SGMs with physics-based methods.
• Applying the coupled framework to multiscale dynamical systems.

Main Results:

• Demonstrated that SGMs can be effectively coupled with physics-based methods.
• Showed improvement in sampling efficiency within multiscale dynamical systems.
• Highlighted the complementary strengths of ML and physics-based approaches.

Conclusions:

• Score-based generative models offer a powerful tool for enhancing sampling in complex systems.
• Coupling SGMs with physics-based techniques mitigates weaknesses and leverages strengths of each approach.
• This integrated framework advances the modeling of multiscale dynamical systems.