Machine Learning-aided Multiscale Modeling of Fatigue Damage in Composite Structures​

AFRL recently completed a project on determining the technical feasibility of new damage tolerance design approaches for composite aircraft structures. However, the average error for blind static strength predictions was 20%, while that for blind fatigue strength predictions was 40%. To realize Air Force’s vision of service life prediction, we need to predict the remaining fatigue life of composite structures with fatigue test data. Cyclically loaded composite structures may fail by matrix fatigue and fatigue delamination. Existing models, however, cannot satisfactorily solve this complex problem: Existing empirical models neglect the microstructural details of composites and significantly sacrifice accuracy for efficiency. Existing multiscale models cannot characterize all failure mechanisms of composites and are mostly for 3D solids, rarely for common structural forms in terms of beams and plates/shells. To meet the critical need for an efficient high-fidelity fatigue model for composite structures, we propose to develop a machine learning-aided multiscale model for fatigue damage in composite structures based on our previous research on fatigue damage model and cohesive zone model (CZM) to handle matrix fatigue and fatigue delamination, respectively. mechanics of structure genome (MSG) to model 3D structures, beams, and plates/shells in a unified manner. Dakota (a general-purpose optimizer) along with a conditional recurrent neural network (RNN) model to accelerate model calibration. We propose the following Phase I objectives: Develop an implicit integration scheme for our CZM for improved accuracy. Develop a conditional RNN model as an efficient high-fidelity substitute for finite element analysis. Develop a Dakota-based calibration tool and calibrate modes I and II delamination parameters. We also propose the following Phase II objectives: Design and conduct fatigue tests on a composite structure and its matrix and interface for blind prediction and calibration, respectively. Develop an implicit integration scheme for our fatigue damage model. Modify Phase I’s RNN model and calibration tool to accommodate matrix fatigue. Develop a self-learning algorithm retraining the RNN with new data during model calibration. Develop an MSG-based multiscale model for fatigue damage in composite structures. Validate the multiscale model by reproducing the structural-level test data and fine-tune the calibrated parameters if necessary. At the completion of this project, we expect to have developed the proposed multiscale fatigue model. The resulting computational tool will achieve unprecedented predictive capabilities in the remaining service life of composite structures. This project will benefit Air Force and related agencies/industries by reducing experiments and iterative adjustments, shortening the design and analysis period, and ultimately cutting down the cost associated with developing and maintaining composite structures.