STELLAR SCIENCE LTD. CO. — Department of Defense STTR Phase I: N23A-T025

Description

Detailed knowledge of the 3D structure of clouds in the atmosphere, and especially forecasting the time evolution of cloud structure and visibility, remains a difficult problem to solve. Stellar Science proposes to apply machine learning (ML) techniques to solve the problem. Our proposed software, the Cloud Casting Neural Network (CloudCastNN), will ingest publicly available satellite data and numerical weather prediction data and output a fully 3D representation of clouds along with a short term forecast of the evolution of the clouds. The ML process will use a convolutional neural network (CNN) enhanced with the relatively new technique of physics-informed neural networks (PINNs). PINNs are neural networks (NNs) that encode physical equations into the training process of the NN, helping the NN converge to the desired solution during training. We will build the CloudCastNN software as a module in Stellar Science’s already existing Varied Interface & Phenomenology Engineering Relationship Suite (VIPERS) ML framework. In conjunction with modeling and forecasting the 3D structure of clouds, we will develop a cloud visualization application. We envision that this application will allow users to fly virtually through the fully 3D cloud field and control their movement and look direction with an interactive user interface. This will enable, for example, the user to visualize the visibility from any point and in any look direction in the system. The final product will include several additional features, such as the ability to input trajectories, export visualizations as movies, overlay other weather data such as wind velocity fields, and generate plots of vertical cloud profiles or other desired features along flight paths.