A theoretical study to clarify the photodesorption process of water on anatase-TiO 2 (101) was pursued in this study using a combined approach based on quantum chemical calculations and quantum dynamical simulations assisted by machine learning of artificial neural networks. First, an embedded cluster model was used to assess three-dimensional potential energy surfaces of ground and excited state of molecular water adsorption. The excited state addressed herein consists of a photogenerated hole which oxidizes the adsorbed water molecule. Next, the approximately 23 000 data points for each surface were fitted using artificial neural networks to generate a dense grid of the data. These fits are a prerequisite for our subsequent quantum dynamical simulations based on the propagation of wavepackets. Eventually, the photodesorption process of water on anatase-TiO 2 (101) was found to consist of a multidimensional mechanism involving a lateral translation of the water molecule toward the bridging oxygen row. Upon relaxation to the ground state, the maximum possible photodesorption probability of 85% was calculated through analyzing specific resonance lifetimes of the chargetransfer state. By comparison with experimental velocity distributions, long resonance lifetimes of over 60 fs can be ascribed to the studied system. Moreover, temperature effects by studying the population of vibronically excited states were included, which turned out to play a minor role in the photodesorption process.