Packed-Ensemble Surrogate Models for Fluid Flow Estimation Arround Airfoil Geometries
This work addresses the computational cost of physical simulations for fluid dynamics applications, though it appears incremental as it builds on existing ensembling methods.
The paper tackled the problem of accelerating fluid flow estimation around airfoil geometries by developing Packed-Ensemble surrogate models, which reduced training time by 25% compared to Deep Ensembles while maintaining performance.
Physical based simulations can be very time and computationally demanding tasks. One way of accelerating these processes is by making use of data-driven surrogate models that learn from existing simulations. Ensembling methods are particularly relevant in this domain as their smoothness properties coincide with the smoothness of physical phenomena. The drawback is that they can remain costly. This research project focused on studying Packed-Ensembles that generalize Deep Ensembles but remain faster to train. Several models have been trained and compared in terms of multiple important metrics. PE(8,4,1) has been identified as the clear winner in this particular task, beating down its Deep Ensemble conterpart while accelerating the training time by 25%.