Deep learning (DL) algorithms are revolutionizing how information is being processed, with multiple applications in the industry and the everyday lives of individuals. This revolution reached scientific research, where such algorithms have started to be applied to handle large amounts of data and to generate more accurate predictions than previous models. At the same time, the black box property of these tools has raised concerns about to which extent they should be used in science. Recently, a new research field called Scientific Machine Learning has emerged, focused on this integration of deep learning in scientific domains. This dissertation explores how scientific models can be improved using deep learning algorithms. Precisely, it comprehends how deep neural networks can improve and complement well-established scientific applications. We approach this problem from an applied research perspective, presenting different studies that employ DL algorithms to improve the capabilities of meteorological and epidemiological models. Our approach contains three studies that developed DL algorithms to handle different types of shortcomings from the original models. Our first study shows that DL algorithms are a powerful tool to improve the output precipitation predictions from meteorological models (which in meteorology is called post-processing) and to generate probabilistic forecasts, and provide relevant insights about its application. The second study uses DL algorithms to increase the spatial resolution and accuracy of the original precipitation maps and analyzes the performance of different models. The resulting algorithms generate realistic precipitation maps that double the resolution of the original forecast. A third study takes a different approach and uses deep neural networks to integrate an additive term into the SIR models’ equations that describe the dynamics of a COVID- 19 outbreak. This additive term learns the incoming force of infection from neighboring regions and improves the ability of the model to capture the right underlying dynamics. One final contribution analyses and recollects the requirements of a workflow concept that enhances the accessibility and reproducibility of DL algorithms in scientific contexts. We found that DL algorithms are powerful tools to complement dynamical systems in different forms, helping to increase their predictive capabilities. The large amount of available data combined with the neural networks’ universal function approximator property makes it a powerful tool to improve existing models. However, when applying DL algorithms to scientific tasks, some considerations need to be observed to enhance and ensure the scientific validity of its predictions.