Thermodynamic properties of aluminum nanoparticles using gaussian approximation potentials

Abstract

We have developed a machine-learned interatomic potential for aluminum nanoparticles with accuracy near density functional theory, using regression-based Gaussian approximation potential. Ten thousand data points from 10 different nanoparticle sizes, ranging from 40 to 123 atoms, are generated to train and validate our potential. Two models have been developed: model A1 exclusively for N=55 nanoparticles and model A2 for a broad range of aluminum nanoparticles. Both models were so trained that the error between the trained and source data in terms of force and energy is minimal. These models were subsequently used to compute the heat capacities and melting temperatures of different aluminum nanoparticles using the multiple histogram technique. Models A1 and A2 demonstrate remarkable accuracy for Al53, Al55, Al60, Al116, and Al128 nanoparticles. Obtained melting temperatures and heat capacities of Al53, Al55, Al60, Al116, and Al128 exhibit excellent agreement with experimental measurements. The melting temperature is ascribed to the phase transition in the nanoparticles analyzed in terms of mean square displacement and Lindemann index. Further both models A1 and A2 have accurately captured all the striking features observed in the experimental results.