The demand for faster development of reliable electronic packages drives the development of virtual qualification approaches. This work introduces the development of neural networks based on data synthesized from finite element calculations. It was the aim to enable predictions of characteristic solder joint strain values and patterns for life time assessment and damage mode identification. For the load scenario of sinusoidal vibration at different temperatures two neural networks were generated. First, with the target to predict equivalent plastic strain values, a RNN model was developed including evaluating the application of dropout layers. The best performing model had 2 hidden layers with 20 neurons per layer and showed very good test performance (MSE =2.29 e-9). Second, a FNN was generated to enable plastic strain pattern predictions. A model with a 3 layer structure and 64 and 128 neurons per layer respectively was found to have good performance (MSE=3.78 e-06) and passed a comparison against finite element calculated strain patterns and damage observations from actual vibration experiments.