Extending length and time scales of first principles material modelling via machine learning
Material modelling based on density functional theory (DFT) nowadays is an indispensable tool in materials research. With increasing computational power and with the advent of data science techniques it could become cheaper and more productive materials discovery laboratory. However, in cases where materials with a large number of atoms in the unit cell are studied and/or when long time dynamics is needed, DFT becomes computationally too expensive. Two examples of such cases of high academic and economic interest are modelling temperature effects in molecular crystals and dynamics of molecules on surfaces. To bridge accuracy of DFT with requirements for long time dynamics with a large number of atoms, machine learning (with artificial neural networks) of potential energy surfaces based on precomputed DFT energies/forces will be employed. Such machine-learned potential energy surfaces enable evaluation of energy/forces with DFT accuracy at a fraction of computational cost. The two most important scientific contributions of this project will be: i) machine learned potentials for molecular crystals that will enable understanding of dynamical and temperature induced phenomena in molecular crystals (as a prominent example a thermosalient effect will be explained), and high-throughput screening of molecular crystals for desirable properties, ii) understanding complex dynamics of molecules on surfaces relevant for heterogeneous catalysis under different excitation, coverage, and pressure conditions through molecular dynamics on machine learned potentials. The ultimate aim of this project is to establish a group that would, with its expertise in DFT reinforced with machine learning techniques, through collaboration with experimentalists, and by educating young researchers, significantly enhance materials research in Croatia.