| Abstract Scope |
In this work, we focus on the lattice configuration problem: given a lattice structure, how does nature prefer to arrange different atomic species under processing conditions? Traditionally, this problem has been treated by Metropolis Monte Carlo simulations using effective Hamiltonians fitted to reproduce first-principles calculations. However, obtaining reliable effective Hamiltonians is often a difficult task for many-component systems of technological relevance. Here, we employ machine learning potentials for this task in an unconventional way: the potential is trained to predict the relaxed energy from ideal on-lattice configurations. The idea is combined with an iterative training approach in combination with extended ensemble Monte Carlo methods to obtain a balanced training set and speed up the sampling. The idea is demonstrated on several spinel oxides as well as solid electrolyte systems. A python software framework abICS for facilitating this process will be introduced. |