Machine-learning Based Screening of Lead-free Halide Double Perovskites for Photovoltaic Applications

Abstract

Lead-free halide double perovskites are promising stable and non-toxic alternatives to methylammonium lead iodide in the field of photovoltaics. In this context, the most commonly used double perovskite is Cs$_2$AgBiBr$_6$, due to its favorable charge transport properties. However, the maximum power conversion efficiency obtained for this material does not exceed 3\%, as a consequence of its wide indirect gap and its intrinsic and extrinsic defects. On the other hand, the materials space that arises from the substitution of different elements in the 4 lattice sites of this structure is large and still mostly unexplored. In this work a neural network is used to predict the band gap of double perovskites from an initial space of 7056 structures and select candidates suitable for visible light absorption. Successive hybrid DFT calculations are used to evaluate the thermodynamic stability, the power conversion efficiency and the effective masses of the selected compounds, and to propose novel potential solar absorbers.