Machine learning potentials struggle to predict silica glass structure

Researchers have investigated the limitations of machine learning potentials in accurately predicting the medium-range order of silica glass. Using neutron and X-ray diffraction alongside molecular dynamics, they found that even models incorporating long-range interactions struggle to replicate the experimental amorphous structure after vitrification. Both short-range and long-range models exhibited excessive ordering and constrained network flexibility, indicating that current approaches are necessary but insufficient for predictive modeling of disordered silica. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Suggests current MLIPs are insufficient for predicting complex material structures, requiring new training data and sampling strategies.

RANK_REASON Academic paper detailing research findings on machine learning potentials for materials science.

Read on arXiv cs.LG →

• arXiv
• Neutron
• X-ray Diffraction
• Machine Learning
• Silica Glass
• MACE
• SiO4

• paper
• other