Physicists have access to Lagrangian ocean models to predict particle densities, where they have certain sets of parameters which they can change to alter the simulation results. The issue with it is that these simulations are fixed. They will not vary every time they run their models, so the question of having uncertainty measurements is not possible. To have some form of uncertainty measurements in the prediction of particle densities, we explore the use of multi-headed Convolutional LSTM (ConvLSTM) for forecasting particle densities and present the network’s confidence in its predictions. We make use of Bayesian approximation techniques to do so. We train and test our network on data generated by Lagrangian ocean simulators. We investigate how different design choices of the network effect the confidence of the model’s predictions. Our experiments show that while the network has struggled with the output accuracy, it represents its shortcomings in its confidence measurements.