Learning mappings of thermal updraft fields under unknown operating conditions using a deep operator network

Precise estimation of the thermal updraft environment is important for the effective exploration of wind resources in long-endurance drones. Nevertheless, previous regression algorithms exhibit limitations in accurately evaluating updrafts under new operating conditions, and traditional airborne wind measurement methods are constrained by narrow ranges and sparse spatial sampling. This study addresses these challenges by harnessing continuous temperature data acquired via infrared sensors. The proposed methodology employs a data-driven deep operator network (DeepONet) to map the temperature field to the velocity field. Numerical simulations of two-dimensional Rayleigh–Bénard convection are conducted to simulate sensing measurements under various Rayleigh number Ra, used as both training and testing datasets. For the DeepONet framework, a convolutional neural network (CNN) structure is employed as the branch network to extract features from the temperature field. Simultaneously, a fully connected neural network (FNN) is adopted as the trunk network, encoding input functions from fixed sensors. In order to assess the estimation performance in new environments, the training data are under operating conditions within the range of Ra=3×107–6×107, and the testing data are under other unknown operating conditions. By compared to the conventional FNN network and the standard DeepONet framework, the DeepONet(CNN) in this study manifests a significant enhancement in estimation performance, demonstrating improvements ranging from 20% to 40% under unknown operating conditions.