Rapid prediction for deflection history of CFRP beams during curing using LSTM network and its application to stacking sequence optimization with genetic algorithm

Predicting process-induced deformation (PID) is crucial for part quality control. However, conventional numerical modeling is inefficient for this task as it requires strict calculation for the entire parts. For improvement, a long-short term memory (LSTM) network was developed to rapidly predict PID of carbon fiber reinforced polymer (CFRP) beams throughout curing. The training database was generated using the finite element modeling (FEM) method with thermo-viscoelastic constitutive law. The principal component analysis, time standardization and logarithm operation were utilized in data pre-processing to enhance prediction accuracy. Afterwards, the LSTM model was integrated with the Genetic Algorithm to optimize stacking sequence of the CFRP beams for minimal PID, with the result experimentally validated to be less than 0.013 mm deviation in final PID of 7.5 cm long samples. Compared to FEM, the LSTM analysis saved 99.9 % of the running time, enabling fast product quality estimation and PID optimization in production lines.