Inverse design of porous structures using deep learning methods for targeted physical properties /

Includes bibliographical references

PhD;2024;Design and Manufacturing

Porous structures have garnered significant attention in various fields of mechanical engineering due to their unique properties and wide range of applications, particularly in thermal, fluid, and structural domains. However, the intricate relationships between the geometry of a porous structure and its target physical properties pose challenges in design. This thesis explores the use of deep learning algorithms to efficiently tackle the inverse design problem of constructing porous structures with user-desired multifunctional properties. The thesis presents three novel deep-learning frameworks to address the inverse design of periodic and irregular porous structures in 2D and 3D. First, a three-stage deep neural network (DNN) framework is proposed to design periodic porous structures with targeted effective thermal conductivities and permeability. The framework introduces a systematic approach to separate geometrical parameters into derivable and controllable categories, enabling accurate geometry prediction based on desired physical properties. Experimental validation is conducted to assess the framework’s performance. Second, a conditional progressive generative adversarial network (ProGAN) combined with a pretrained property predictor is developed for the inverse design of irregular porous structures simultaneously satisfying multiple properties, such as porosity, transverse elastic moduli, and thermal conductivity. This approach works with pixel representation of the structure (images). This capability is important as it captures the complex geometry of irregular structures. The framework leverages a large dataset of 1,00,000 2D periodic representative volume element (RVE) images and their corresponding effective properties obtained through high-fidelity finite element simulations. A novel conditions input mechanism is introduced to handle multiple target properties during the training process. Third, the conditional generative adversarial networks (CGANs) and multiple condition vinput mechanisms are extended to develop a framework for generating 3D Voronoi-based porous lattice structures. This is an extension of the above approach to 3D, where the data sets are now voxel-based. A dataset containing 20,990 3D CAD models (represented by voxels) and their corresponding physical properties is constructed, and a 3D conditional GAN is trained on this dataset. The trained model successfully generates novel 3D lattice structures that meet user-specified elastic and thermal properties in the x, y, and z directions. The proposed deep learning frameworks demonstrate the efficacy of data-driven approaches in tackling inverse design problems for porous structures. The frameworks offer significant advancements over traditional design methods, providing faster and more efficient ways to generate porous structures with targeted multi-functional properties. The methods highlight the potential of deep learning in accelerating the inverse design process. The frameworks developed herein can be extended to incorporate additional physical properties and adapted to other porous structures.