Machine-learning investigation of layer-resolved collective dynamics in interfacial liquid Fe

Understanding how local symmetry breaking at interfaces affects the collective dynamics of liquids remains a long-standing challenge in condensed matter physics. In this work, molecular dynamics (MD) simulations combined with the intrinsic sampling method (ISM) were used to obtain layer-resolved structural and dynamical information for molten Fe at its free surface. Based on these data, shallow neural-network (SNN) models were trained to establish quantitative mappings between the intermediate scattering functions (ISFs) of the bulk liquid and those of interfacial layers. The models successfully reproduce the crossover from surface-induced slowing at small wavevectors to dynamic acceleration at large wavevectors, reflecting the spatial complexity of collective relaxation in the interfacial region. Cross-temperature validation shows that the trained networks retain reasonable predictive capability for short-wavelength modes but fail to describe long-wavelength collective motions, emphasizing the need for physically informed features such as local order or interatomic correlations. Overall, this work provides a first step toward a data-driven approach to interfacial liquid dynamics, offering a scalable framework to interpret and predict collective behaviors of liquids across different interfaces and thermodynamic states.