Speaker
Description
The DANTE project aims to create computationally efficient models of urban microclimate by applying model order reduction techniques to high-resolution urban-scale simulations. Resulting models must undergo a rigorous validation process before any application is possible, to ensure their accuracy and quantify their uncertainties. This validation process requires urban-scale ground truth data, which is not directly available. Instead, lower-resolution data must be downscaled to urban scale.
The lack of available data and models prevent the use of typical statistical and dynamical downscaling methods. Furthermore, the inhomogeneity of the scales of relevant flow structures requires that both input and target resolution data are irregular meshes, rather than grids. For these reasons, traditional downscaling methods are unsuitable. The goal of our work is to construct a downscaling framework adapted to the context of weather downscaling, leveraging regional model data, weather station measurements, and physical knowledge.
One solution is Physics-Informed Neural Networks (PINNs). PINNs incorporate physical constraints into the learning process by including PDE residuals into the loss function. By using a network that takes coordinates as input and outputs the local system state, a fitted model can be evaluated at arbitrary coordinates, providing a way to downscale (continuous PINN). However, a major downside of PINN is their lack of robustness: it can be difficult to get them to reliably converge.
In this context, we explore the difficulties that come from using PINNs, and ways around them. We define a criteria for PINN convergence based on the influence of the inclusion of physics in the loss, and study the influence on the PINN training process of architecture, collocation point density, weighting scheme of loss terms, preprocessing, and training protocol. We also present limitations of this method that we are not yet able to overcome, such as the dependency to initialization.