Session: 12-05-02: Data-Enabled Predictive Modeling, Scientific Machine Learning, and Uncertainty Quantification in Computational Mechanics

Paper Number: 99024

99024 - Deep Learning Accelerated Topology Optimization With Inherent Control of Image Quality 

Deep learning (DL) is a powerful tool to accelerate topology optimization toward a wide range of engineering applications that demand instantaneous conceptual design with high accuracy and precision. However, many existing DL-based topology optimization methods predict structures that have low image quality with significant blur and distortions, obstructing direct manufacturing of the designed parts. To address the technical challenge, this study proposes a DL model based on the deep Residual U-net (ResUnet) architecture, a convolutional neural network (CNN) that is efficient and accurate even without a large amount of data. A new loss function is proposed for DL-based topology optimization by combining the complementary distance-based and similarity-based loss functions. It has two parameters that are optimized to achieve the best performance considering four criteria, i.e., maximum image quality, minimum compliance error, minimum volume fraction error, and minimum structural discontinuity. Trained with the optimal structures under a variety of loading and boundary conditions, the present DL model can predict optimized structures almost instantaneously, with high image quality readily for manufacturing. The image quality improvement, along with the other performance improvements, is shown to persist regardless of the training image quality or the resolution of the problem. The highly universal loss function is expected to provide an inherent approach to improve the manufacturability of DL-predicted structures and extend the application of DL-based topology optimization methods. In summary, a DL model is proposed based on ResUnet for instantaneous topology optimization that features an inherent ability to improve image quality for direct manufacturability. The NN is trained with datasets that include (1) an input feature set that describes the design problem including the computational domain, loadings, boundary conditions, and materials; and (2) the optimized structure for the described problem, which can be generated by any topology optimization methods, e.g., SIMP for this study. Several input feature sets are compared and the optimal one that balances efficiency and accuracy is found to include the stress components, the von-Mises stress, and the magnitude of displacement, all evaluated pointwise on a uniformly “grey” plate with the specified volume fraction. With this optimal input feature set, the performance of the DL model quickly converges with increasing mesh size or the increasing resolution of the problem. The combined loss function has shown excellent performance regardless of the training image quality. When the image quality is low, the combined loss function improves image quality, while the other loss functions keep the low image quality almost unchanged; and when the image quality is high, the combined loss function retains the high image quality, while the others substantially reduce image quality. In any case, the new loss function predicts structures that have high image quality along with suppressed discontinuity and well-controlled compliance error. The new proposed loss function has shown amplified performance improvement when the problem has low resolutions.

Presenting Author: MD MOHAIMINUL ISLAM Temple University

Presenting Author Biography: Md Mohaiminul Islam is currently a Ph.D. student at Temple University in Mechanical Engineering department. He is a Loretta C. Duckworth Scholar Studio extern fellow of the Temple University library. His research interest lies in interfacial thermal conductance, multiscale modeling, and data-driven predictive modeling.

Authors:

MD MOHAIMINUL ISLAM Temple University
Ling Liu Temple University