Halide perovskites have emerged as a promising class of materials for photovoltaic applications, such as solar cells. In order to tune the relevant opto-electronic properties of these materials one commonly introduces mixing and alloying, for example by mixing CsPbI3 and CsPbBr3 leading to a the CsPb(Br,I)3 alloy. Importantly alloying can also lead to stronger resistance against the system degrading to photoinactive phases, which is a common problem in these materials. Here, we aim to investigate the details of mixed halide perovskite on a atomic-scale using atomic scale models for the prototypical system CsPb(Br,I)3. To this end, we will train neural network models using density functional theory (DFT) calculations (carried out elsewhere) as training data. In these mixed systems we have to consider both vibrational and configurational degrees of freedom, and hence the input space to the neural network is very large. Therefore, in order to generate suitable training structures we will use an active-learning scheme to address this problem. This approach relies on computing the model uncertainty for new samples. If the uncertainty is above a suitably chosen threshol, the sample will be included in the training set for the next model generation. In order to evaluate the uncertainty of our models we will train a committee of models. Furthermore, the sampling of both vibrational and configurational degrees of freedom can be computationally expensive since it requires many thousands or even millions of evaluations of the underlying model. Here, the fact that our models are implemented on GPUs makes it incredibly fast to evaluate, which is a crucial advantage of our approach.