Life Prediction for FRP composites with Data Fusion & Machine Learning

High-fidelity, probabilistic predictions of damage evolution in fiber-reinforced polymer (FRP) composite structures could accelerate development and certification of new concepts through a reduction in the need for physical testing and an increase in structural health awareness. Unfortunately, high-fidelity models often carry the burden of excessive computation times. Probabilistic damage prognosis involves fusing structural health data with physics-based models using what are sometimes referred to as inverse uncertainty quantification techniques. Here, model parameters that cannot be determined directly are calibrated based on measurement data, and the associated parameter uncertainties are simultaneously quantified. However, these techniques require thousands or even millions of simulations to converge, which compounds the computational problem associated with high-fidelity models. The current state-of-the-art in damage prognosis avoids this issue through a simplification of the damage models, which in turn sacrifices the potential of these predictions to significantly impact research and development at NASA. This proposal aims to achieve a high-fidelity, probabilistic predictive capability for FRP composites utilizing data fusion, advanced modeling and simulation, and machine learning.

Objective: Reach beyond the state-of-the-art utilizing machine learning and data fusion to enable probabilistic life predictions for real-world structures comprised of fiber-reinforced polymer (FRP) composite materials.

Problem: Accurately predicting the remaining life of these structures requires the integration of high-fidelity damage progression models, uncertainty quantification (UQ), and inverse uncertainty quantification (iUQ) for data fusion (e.g., Markov chain Monte Carlo methods). Unfortunately, state-of-the-art, high-fidelity damage models can be excessively time consuming which often prohibits their use with standard iUQ methods (e.g., a predictive analysis with 10,000 simulations, each with a runtime of ~3 hours, would take on the order of years to complete). The issue is further complicated in that parallelization of iUQ techniques is inherently restricted.

Current Practice: Probabilistic life predictions are typically conducted using simplified one- or two-dimensional analytical models that are time-efficient for use with iUQ. However, these models sacrifice fidelity and are generally incapable of representing damage evolution in real-world structures that are complex in geometry, loading, and material. Damage in these structures is subject to mixed-mode driving forces and progresses in three-dimensions.

Innovations: (1) The aforementioned issue of prohibitive computation times will be addressed using surrogate models in place of time consuming finite element analyses (FEA). Structural health monitoring data will then be fused with the proposed high-fidelity damage modeling framework to form probabilistic life predictions for real-world structures using iUQ. Based on initial estimates, prediction times have the potential to be reduced by almost three orders of magnitude using this approach. (2) To accomplish the above goal, state-of-the-art composites damage progression research (see attached supplemental material) will be implemented in the three-dimensional fracture mechanics software FRANC3D, developed by Fracture Analysis Consultants, Inc. Machine learning techniques will be used to train the surrogate models utilizing FRANC3D simulations completed a priori and in parallel, leveraging high performance computing.