SMARTy
The SMARTy package, written in pyhton provides methods for various data-driven tasks such as predictive modeling of scalar or high-dimensional quantities. For predicting scalar quantities various state-of-the-art interpolation and regression models are available such as Gaussian Processes. Established techniques for dimensionality reduction (DR) like proper orthogonal decomposition are implemented together with machine learning alternatives such as Isomap or Autoencoder. Reduced order models (ROMs) can be derived by combining DR methods and regression or intrusive approaches to predict high-dimensional quantities. Moreover, neural networks can be constructed utilizing established deep-learning libraries such as TensorFlow or PyTorch with a unified wrapper. SMARTy comes along with Bayesian optimisation techniques for hyperparameter optimisation. For interpolation and regression models, ROMs as well as neural networks a common user-interface layer is used to learn the input-output mapping of the given data. If no data is available or new data is required, SMARTy provides design of experiment methods to gather data at optimal sampling locations. Gradient-based as well as gradient-free optimisation is possible either relying on surrogates or using other established algorithms. Moreover, also multi-objective with and without constraints are possible.
Your consent to the storage of data ('cookies') is required for the playback of this video on Quickchannel.com. You can view and change your current data storage settings at any time under privacy.
Uncertainties can be introduced and efficiently propagated though black-box functions for tasks like robust design. Data from different sources, e.g. numerical simulation and experimental results, can be combined into one model either using multifidelity modelling or data fusion approaches such as gappy POD/Isomap. During this process uncertainties present in the data can be accounted for. As a FlowSimulator Plugin SMARTy provides machine learning approaches for high-fidelity numerical simulation like data-driven turbulence modeling. In addition, it can be easily integrated into complex workflows for multidisciplinary optimisation.
Key features
- surrogate modeling and multi-fidelity surrogate modeling
- design of experiment methods
- intrusive and non-intrusive reduced order modeling
- machine learning for aerodynamic applications
- automatic model evaluation and selection
- data fusion
- surrogate-based as well as gradient-based optimisation
- surrogate based uncertainty quantification
- robust design
- FlowSimulator plugin with connection to CFD-solvers (e.g. TAU, CODA)
Areas of application
SMARTy is used for aerodynamic modeling, analysis and design. In particular, it focuses on the fast prediction of steady aerodynamic data such as performance metrics and loads throughout the entire flight envelope. Furthermore, SMARTy is also applied for the efficient prediction and analysis of unsteady aerodynamic responses. Another application area of the software is the combination of classical CFD-methods with machine learning methods, e.g. for data-driven turbulence modeling. Furthermore, SMARTy provides models for multi-disciplinary analysis and optimisation and overall aircraft design. In addition, SMARTy is used for the quantification and propagation of uncertainties through complex black-box models as well as for robust optimisation. The fusion of numerical and experimental data is also possible using SMARTy.
Links
License notice
Access to these software products can only be granted if the organisation you are representing holds a valid license or your organisation is interested in signing such a license. If either of the aforementioned two points are fulfilled please get in touch with us providing additional information. Afterwards, we will support you in getting access to our software products. Please note, that we are, in general, not signing license agreements with private individuals.