SUMMARY

Deep learning has developed rapidly in the past decade and has found applications in many fields. Neural networks are sometimes referred to as “universal function approximators” for their ability to learn complex problems that are otherwise difficult to solve. In engineering, there are many complex problems that can be solved by neural networks. Neural networks are also computationally efficient and can be applied to time- consuming problems, such as those that require simulation software or optimization algorithms. In this research, convolutional neural networks (CNN) will be trained to predict 2D stress responses in solid cantilever beams subjected to loading, as well as thermal responses in 2D heat transfer fins. The datasets used in these problems exhibit highly skewed distributions of values, which is called data imbalance and can limit the accuracy of the networks. Two methods for addressing data imbalance and improving prediction accuracy will be demonstrated: a data sampling method, and a label distribution transformation method. In addition, a CNN will be trained to design 3D lattice structures based on relative density information.