Deep learning-based predictions of process parameters for thick composites laminates manufacturing

Ammasai Sengodan, Ganapathi; Hoang Luong, Minh; A Ponnusami, Sathiskumar

Abstract

In a composite manufacturing process, the tool geometry and thermal characteristics of the bagging materials dictate the heat transfer mechanism. The temperature distribution, resin chemical kinetics and residual thermal strains determine the cured part quality. These mechanisms lead to manufacturing anomalies and impede the design for manufacturability and assembly. The exothermic heat generated at the centre of the thick laminates results in thermal overshoot and affects the resin re-distribution. The uneven resin distribution leads to residual deformation and part distortion, affecting the final shape of the composite part. The degree of cure and temperature profile at the centre of the curing laminate is determined numerically or experimentally to optimise the prepreg manufacturer given cure cycles. High fidelity finite element process modelling simulations aid composite engineers in designing and manufacturing decisions that produce high-quality laminates. The rapid growth of artificial intelligence and data science techniques could minimise the computational efforts by learning the relationship between the input process parameters and the pre-defined outcomes.

In this work, a multi-layer perceptron (MLP) was built to learn the relationships between the input parameters such as tool & part geometry, heat transfer coefficients and the resulting degree of cure and temperature profiles. The data set required to train the MLP was generated from the thermo-chemical process simulation using ABAQUS/Explicit with user-defined subroutines. The MLP was constructed to learn the data of three different prepreg material systems such as AS4/8552, AS4/3501-06 and a Glass Polyester. The results suggest that the MLP model can predict the degree of cure (DOC) and temperature profile with an accuracy of 99.8 % and 91.8 %, respectively. Thus, a deep neural network is trained to anticipate the temperature and degree of cure profiles of the laminates for input parameters. In future, the MLP will be extended to predict and optimise the part distortion, such as spring-back, corner thickening and edge deformations.

Citation

Ammasai Sengodan, G., Hoang Luong, M., & A Ponnusami, S. (2022, September). Deep learning-based predictions of process parameters for thick composites laminates manufacturing

• Presentation Conference Type: Speech
• Conference Location: Sheffield, United Kingdom
• Start Date: Sep 14, 2022
• Deposit Date: Sep 8, 2023