60158 - Using Autoencoders and Output Consolidation to Improve Machine Learning Models for Turbomachinery Applications 

Machine learning models are becoming an increasingly popular way to exploit data from fluid dynamics simulations. Here we investigate how autoencoders and output consolidation can be used to increase the accuracy of models, by injecting knowledge of the full flow field in the predictions of Quantities of Interest (QoI) for optimisation.

The work is carried out on the Workshop platform by Monolith [1], a machine learning software specifically conceived for engineering applications, using simulation data provided by Rolls Royce for a highly loaded transonic rotor.

Autoencoders are machine learning models which are trained to parametrise a given 3D geometry from a set of 3D surface mesh files. This means that an accurate autoencoder can be obtained without the need to run expensive flow simulations. The values of the identified parameters can then be recovered by the model for any unseen mesh with the same structure. This makes autoencoders a key complement to scalar data, as they provide a more accurate description of the geometry and can enhance the accuracy of the models which predict key flow quantities. Once the optimal parameters have been found, the autoencoder is able to reconstruct the corresponding 3D geometry. The ability to easily visualise the output makes the process more intuitive and interpretable.

Quantities of Interest describing the performance of a turbomachine are calculated from relevant simulation outputs, such as pressure fields or velocity profiles. Output consolidation is a method which allows to incorporate multiple relevant datasets into a single specific prediction. It uses two models in succession, in such a way that: the primary model predicts an intermediate quantity (related to the QoI) from the given scalar or geometry data; the secondary model predicts the QoI by taking as inputs both the given data and the primary model predictions. Modelling the intermediate quantities which are normally used to calculate the QoI allows us to produce a secondary model which is informed about the flow. This connection which is created between the geometry data and the field quantities yields an improvement in the predictions of the QoI, as the intermediate data guides the secondary model towards a more “physically sensible” prediction.

Baseline predictions are first generated using standard methods such a kriging and the performance of models based on autoencoders and output consolidation is compared to this benchmark. The hyperparameters for each model are optimised by grid search or by comparing different combinations of kernels. Results show promising improvements in the accuracy of the predicted QoI. Using an autoencoder, it is possible to train models on 14% fewer expensive flow simulations, while also predicting changes in the QoI with a 10% increase in accuracy. A combination of the two techniques described above produces further improvements, and the higher accuracy of the models potentially leads to more insightful optimisation studies. This is supported by the fact that the prediction accuracy notably increases in the range of QoI values which is involved in optimisation problems, making the presented methods particularly attractive.

Ref

[1] More information can be found at: https://www.monolithai.com