In this study, we identify the local structures of ex-solved nanoparticles using machine-learned potentials (MLPs). We develop a method for training machine-learned potentials by sampling local structures of heterointerface configurations as a training set with its efficacy tested on the Ni/MgO system, illustrating that the error in interface energy is only 0.004 eV/Å2. Using the developed scheme, we train an MLP for the Ni/La0.5Ca0.5TiO3 ex-solution system and identify the local structures for both exo- and endo-type particles. The established model aligns well with the experimental observations, accurately predicting a nucleation size of 0.45 nm. Lastly, the density functional theory calculations on the established atomistic model verify that the kinetic barrier for the dry reforming of methane are substantially reduced by 0.49 eV on the ex-solved catalysts compared to that on the impregnated catalysts. Our findings offer insights into the local structures, growth mechanisms, and underlying origin of the catalytic properties of ex-solved nanoparticles.
Keywords: dry reforming of methane; ex-solution; machine learning; machine-learned potential; nanoparticle.