Convolutional neural networks for advanced adhesive joints application patterns

Abstract Adhesive bonding is a widely used joining technique across various industries. Achieving uniform adhesive coverage over the entire surface without the formation of air pockets is crucial for creating strong and durable joints. Simultaneously, it is essential to minimise waste caused by material leakage at the edges. However, generating an optimal adhesive pattern to achieve the desired adhesive distribution after compression remains a challenge, as fluids tend to spread in a circular manner, while industry-relevant target geometries are typically non-circular. This paper investigates the application of Convolutional Neural Networks (CNNs) to optimise adhesive application patterns by utilising a simplified simulation model known as the Partially Filled Gaps Model (PFGM) to generate extensive training data. The CNN is trained to predict fluid distribution outcomes based on initial adhesive application patterns and addresses the inverse problem of determining an optimal application pattern to achieve a desired target distribution after compression. Two training approaches are introduced: a basic inverse model that utilizes a straightforward input–output data exchange, and a more advanced strategy that incorporates a forward model to improve accuracy. The forward model predicts the final distribution, enabling better refinement of the initial application patterns. The results demonstrate that the CNN-based approach is highly effective in generating optimal application patterns for adhesive bonds. Its primary advantage, compared to alternative methods, lies in its ability to achieve precise results within a short computation time. However, a significant drawback is the limited flexibility in accommodating variations in parameters.