Exploring the landscape of Buckingham potentials for silica by machine learning: Soft vs hard interatomic forcefields

Interatomic forcefields for silicate glasses often rely on partial (rather than formal) charges to describe the Coulombic interactions between ions. Such forcefields can be classified as “soft” or “hard” based on the value of the partial charge attributed to Si atoms, wherein softer forcefields rely on smaller partial charges. Here, we use machine learning to efficiently explore the “landscape” of Buckingham forcefields for silica, that is, the evolution of the overall forcefield accuracy as a function of the forcefield parameters. Interestingly, we find that soft and hard forcefields correspond to two distinct, yet competitive local minima in this landscape. By analyzing the structure of the silica configurations predicted by soft and hard forcefields, we show that although soft and hard potentials offer competitive accuracy in describing the short-range order structure, soft potentials feature a higher ability to describe the medium-range order.Interatomic forcefields for silicate glasses often rely on partial (rather than formal) charges to describe the Coulombic interactions between ions. Such forcefields can be classified as “soft” or “hard” based on the value of the partial charge attributed to Si atoms, wherein softer forcefields rely on smaller partial charges. Here, we use machine learning to efficiently explore the “landscape” of Buckingham forcefields for silica, that is, the evolution of the overall forcefield accuracy as a function of the forcefield parameters. Interestingly, we find that soft and hard forcefields correspond to two distinct, yet competitive local minima in this landscape. By analyzing the structure of the silica configurations predicted by soft and hard forcefields, we show that although soft and hard potentials offer competitive accuracy in describing the short-range order structure, soft potentials feature a higher ability to describe the medium-range order.