Learning a reactive potential for silica-water through uncertainty attribution.

Area of Science:

• Materials Science
• Computational Chemistry
• Geochemistry

Background:

• Silicate reactivity in aqueous solutions is crucial for geology, cement chemistry, catalysis, and materials science.
• Accurate molecular-level simulations of these systems are computationally challenging due to scale and accuracy requirements.

Purpose of the Study:

• To develop and validate a machine-learning reactive interatomic potential for silicate-water interactions.
• To demonstrate the potential's ability to capture complex reactivity and properties.

Main Methods:

• Training a machine-learning potential based on the PaiNN architecture using a large dataset of energies and forces.
• Implementing an active learning strategy to iteratively improve model accuracy by focusing on uncertain regions.
• Utilizing enhanced sampling simulations to study specific reactions like water self-ionization and silicate dimerization.

Main Results:

• The developed potential accurately reproduces static and dynamic properties of water and silicates, despite training on cluster data.
• The active learning strategy effectively improved model robustness for bulk simulations.
• Enhanced sampling simulations accurately captured water self-ionization and silicate oligomer acidity.
• Silicate dimerization in water was found to proceed via a flanking mechanism.

Conclusions:

• Machine learning, specifically the PaiNN architecture with an active learning strategy, offers a powerful approach for simulating silicate-water reactivity.
• The developed potential provides a computationally efficient and accurate tool for studying diverse silicate-based chemical systems.
• This work advances the simulation capabilities for complex aqueous silicate chemistry relevant to multiple scientific disciplines.