Image-based topology optimization using a deep learning-driven optimizer network

Abstract

To tackle the challenges of high computational costs, complex parameter tuning, and convergence, this paper introduces the Optimizer-Net model. This groundbreaking deep neural network leverages image-based datasets to supersede explicit programming and lengthy, time-consuming, and iterative numerical computations, seamlessly integrating computer vision into topology optimization. We developed a benchmark method for generating image datasets based on structural behavior. The dataset comprises pairs of images: energy contour images derived from normalized energy, representing structural behavior under applied loads, and the corresponding optimized structure images to showcase diverse features, including textures, colors, and contour variations, providing a rich foundation for training the model. Optimizer-Net analyzes the high-dimensional information of energy contour data and extracts latent features from the images, utilizing optimized structures as a mask to effectively guide the training process. The model was evaluated using two loss functions: Mean Squared Error (MSE) and Cross-Entropy. Results show consistently decreasing training and validation losses, demonstrating superior optimization performance, with MSE achieving 97.123 % accuracy in predicting optimal structures. Optimization times were improved significantly, reducing to 0.219 seconds with MSE and 0.244 seconds with Cross-Entropy. By circumventing typical constraints like mesh grids, iterative loops, and computationally intensive analyses, Optimizer-Net enhances process efficiency, delivering near-instantaneous results.