Transfer learning-based layout inverse design of composite plates for anticipated thermo-mechanical field

The thermo-mechanical performance of composite system can be improved by layout design, which is a hot topic in many engineering fields such as aerospace, electronic systems, etc. Since the design space is complex and enormous, traditional design methods such as simulation and optimization algorithm suffer from time consuming and high cost. To overcome these difficulties, we propose a lightweight deep learning based inverse design method (DLBIDM) which can directly generate qualified layouts from given thermo-mechanical fields. In addition, we use transfer learning to apply the DLBIDM to different thermal conditions, material volume fraction, and larger scale. Inverse designing a layout takes an average of 0.0284 s with almost no loss of accuracy, which greatly improves design efficiency. Compared to existing literature, the model requires a training dataset size that has decreased by two orders of magnitude. This study provides an effective solution for the layout inverse design of composite system from properties to its structures. © 2024 Elsevier Ltd