Deep learning allows overparameterized neural networks with multiple layers to succes- sively extract higher-level features from the raw input. These networks are well known to deal with supervised learning tasks, which require a large amount of labelled training data. To avoid data collection, which is normally expensive in engineering applications, it is critical to use the method with less data dependency and train deep learning mod- els using primarily constraints (physical laws) rather than data. Physics-informed neural networks (PINNs) can be a potential solution [1]. PINNs are capable of leveraging the underlying laws of physics to extract patterns from high-dimensional data generated from experiments. We present deep learning phase-field models for brittle fracture [2]. A variety of physics- informed neural networks (PINNs) techniques, for example, original PINNs, variational PINNs (VPINNs), and variational energy PINNs (VE-PINNs) are utilized to solve brittle phase-field problems. The performance of the different versions is investigated in detail. Also, different ways of imposing boundary conditions are examined and are compared with a self-adaptive PINNs approach in terms of computational cost. Furthermore, the data- driven discovery of the phase-field length scale is examined. Finally, several numerical experiments are conducted to assess the accuracy and the limitations of the discussed deep learning schemes for crack propagation in two dimensions. We show that results can be highly sensitive to parameter choices within the neural network.