Physics-Informed Machine Learning Model for In-situ Life-Cycle Prediction of Condensation Trails

We develop the first end-to-end machine learning (ML) pipeline for prediction of the life cycle of condensation trails, i.e., starting with engine operating conditions leading to contrail formation, followed by cirrus cloud persistence. The few existing contrail models use simplified physics-based models, and therefore the predications deviate from the observations by extending the prediction horizon or in cases where the size of the contrail extends dramatically. Our pipeline is comprised of three stages: (i) given the flight and weather condition an auto-encoder-based regression predicts the profile of the initial phase of contrail formation of turbofan engines, i.e., double-jet mixing. (ii) A similar architecture then predicts the short-term horizon of formation, where the contrail substantially has evolved from its initial condition; (iii) a long-horizon LSTM is then rolled-out model with weather condition as exogenous forcing, using the output of the previous stage model as the first few time-steps of prediction. In our developed framework, high-fidelity physics-based simulations are used to generate the data for training. We only off-load the complicated physics of cloud microphysics to the ML models and retain the first-principal model of convection. At the inference, the model queries “an online” weather database to emulate an in-situ application of the model.