DDPS | “AutoEncoders for Aerodynamic Predictions”
HTML-код
- Опубликовано: 11 сен 2024
- • DDPS Talk date: June 21, 2024
• Speaker: Gianluca Iaccarino (Stanford University, engineering.st...)
• Description: What is an autoencoder? How does it work? How can one trust its predictions? The talk will focus on recent activities centered around the development of an autoencoder, an unsupervised data-driven model, to predict the flow past wing geometries. The model relies on non-linear compression to construct a low-dimensional latent representation of the available data and its relation to the physical inputs. This enables the approach to generate new (unseen) cases. A careful construction of the dataset produces latent variables that can be interpreted in terms of aerodynamic performance both for attached and separated flow conditions. An important thrust of the work is the investigation of effect of uncertainties due to the autoencoder architecture, the hyperparameters and the amount of the training data (internal or model-form uncertainties). Comparisons to a Gaussian Process regression and linear compression strategies illustrate the advantage of the present approach in extracting useful information on the prediction uncertainty even in the absence of data. The effect of model (internal) uncertainties is also compared to the impact of the variability induced by uncertain operating conditions (external uncertainties) showing the importance of accounting for the total uncertainty when establishing prediction confidence. A brief discussion of how to incorporate multi-fidelity data in the autoencoder training will conclude the presentation.
• Bio: Gianluca Iaccarino is the Director of the Institute for Computational Mathematical Engineering and a professor in the Mechanical Engineering Department at Stanford University. He received his PhD in Italy from the Politecnico di Bari (Italy) before joining the faculty at Stanford in 2007. Since 2014, he is the Director of the PSAAP Center at Stanford, funded by the US Department of Energy focused on multi-physics simulations, uncertainty quantification and exa-scale computing. He received the Presidential Early Career Award for Scientists and Engineers (PECASE) award and he is a Fellow of APS.
• Q&A session questions:
a. How does the method handle noise in the input training or testing data, perhaps due to measurement error in sensors?
b. How does the computational cost of the autoencoder compare to that of a RANS fluid simulation of flow over an airfoil?
c. Thank you very much for this great presentation! How can you predict two different values for the same angle of attack values for noise up and noise down motions of dynamic stall from the snapshot trained machine learning model?
d. I'm just learning about these ML techniques, but from what I have seen Variational Autoencoders are more ideal for generation. Does a basic AutoEncoder work well here due to the more narrow/specific scope of the problem? Thanks for the presentation!
e. Can you please explain a little about sequential training for the multi fedelity problems - Do we train the low fidelity's data with low latent and then increase the problem complexity by introducing viscosity etc and increase the latent to see how the new physics is handled?
f. With your method of multi fidelity data, for noisy experiment data for instance, could you add a new latent parameter, but have it be a variational latent parameter and keep the rest as their fixed deterministic latents?
g. The plots shown are CD/CL polars, if the integrated quantity is the goal, would the training benefit or worsen from using the error in the coefficient predictions in the loss rather than just the MSE loss?
h. What differences does the training process have for 3D, unsteady, turbulent flow?
DDPS webinar: www.librom.net...
💻 LLNL News: www.llnl.gov/news
📲 Instagram: / livermore_lab
🤳 Facebook: / livermore.lab
🐤 Twitter: / livermore_lab
About LLNL: Lawrence Livermore National Laboratory has a mission of strengthening the United States’ security through development and application of world-class science and technology to: 1) enhance the nation’s defense, 2) reduce the global threat from terrorism and weapons of mass destruction, and 3) respond with vision, quality, integrity and technical excellence to scientific issues of national importance. Learn more about LLNL: www.llnl.gov/.
IM release number is: LLNL-VIDEO-865917