Active learning of Neural Network proxies for aerodynamic optimization


Many practical continuous minimization problems, such as aerodynamic optimization, are not amenable to gradient-based optimization methods because derivatives can not be computed directly. We recently showed that it is possible to train a Neural Network regressor as a proxy to the numerical simulator and optimise the proxy function via Gradient-Descent. [ICML 2018 Paper] However, in practice, each datapoint requires an expensive call to a numerical simulator. Therefore, we need to use active learning techniques or reinforcement learning-based strategies to optimally sample the initial dataset and re-use the simulator judiciously during the optimization phase.

Shape Optimization with GCNN.


An efficient active learning strategy to take the best advantage of a limited number of calls to the simulator before and during the optimization process would have a strong theoretical and industrial impact.

Starting point

The student will be implementing his work on top of a large state-of-the-art code base, which includes a Deep-Learning library for 3D shapes in the TensorFlow framework.

Key Domains

Please send an email to or

The work will be done in collaboration with the EPFL startup Neural Concept .