Neural Operators: Advancing Real-Time Structural Response

About the Webinar

Traditional numerical modeling techniques, while accurate, often prove limiting for real-time predictions due to their computational intensity. Machine learning (ML)-based surrogate models have emerged as a faster alternative. However, these surrogates are constrained by their training data. For instance, an ML model trained to predict structural displacements based on specific earthquake (EQ) records is limited to that finite dataset, as it fails to capture the underlying relationship between earthquake loading (input functions) and displacement (output functions).

Neural Operators address this limitation by learning function-to-function mappings. A notable example in this category is the Deep Operator Network (DeepONet). The DeepONet architecture comprises two neural networks: a branch net that processes the input function, and a trunk net that handles the locations where the output function is evaluated. This structure enables DeepONet to learn the mapping between entire function spaces, rather than just specific input-output pairs.

In the context of structural analysis, DeepONet can potentially learn the general relationship between loading conditions and the resulting deflected shapes of beams. This approach promises greater flexibility and generalization capability compared to traditional analytical methods, that needs to be simulated for every loading condition. In the webinar, we will work through live coding examples exploring the application of DeepONet in structural engineering. We will demonstrate the development of a data-driven predictive model to analyze the deflected shape of cantilever beam under various loading conditions.

Presenter

Somdatta Goswami is an Assistant Professor at the Department of Civil and Systems and holds a joint appointment with the Department of Applied Mathematics and Statistics. She is also an affiliate member at Institute for Data Intensive Engineering and Science and the Hopkins Extreme Materials Institute at the Johns Hopkins University. Somdatta received her PhD in Engineering at Bauhaus University-Weimar, Germany in phase field-based fracture modeling. Her research interests span Scientific Computing, Computational Mechanics, and Machine Learning, focusing on both fundamental and applied aspects. Her work addresses long-time horizon problems and challenges in multi-scale, multi-physics material modeling. Dr. Goswami is particularly dedicated to developing AI-accelerated numerical simulations, aiming to enhance the efficiency and accuracy of computational processes in materials science and engineering.