A Theory-guided Probabilistic Machine Learning Framework for Accelerated Prediction of Process-induced Deformations in Advanced Composites
Topic(s) : Special Sessions
Co-authors :
Caleb SCHOENHOLZ (UNITED STATES), Enrico ZAPPINO , Marco PETROLO (ITALY), Navid ZOBEIRY (UNITED STATES)Abstract :
While carbon fiber-reinforced polymer (CFRP) composites have seen widespread use throughout the aerospace industry, manufacturers continue to face several challenges. One such challenge is the prediction and control of process-induced deformations (PIDs) in composite parts. During processing (e.g., autoclave), residual stresses form due to complex phenomena in the material and manufacturing environment. Upon demolding, some of these stresses may be released through deformations such as angle alterations at geometry transitions (e.g., spring-in) or warping of initially flat sections. Consequently, these PIDs may induce joining gaps during assembly, prolong production timelines, and compromise the final structure’s mechanical efficiency.
Despite having a general understanding of PIDs, many manufacturers often struggle to accurately predict deformations in industrial settings. These difficulties primarily stem from limitations of traditional methods used for PID predictions, with one of the most notable constraints being the trade-off between fidelity/accuracy and time/cost. This trade-off frequently presents a dilemma for manufacturers when choosing a suitable prediction method, leading to attempts involving multiple methods and inadequate data. As a result, manufacturers are left with inaccurate predictions and seemingly uncontrollable PIDs.
Given challenges previously outlined, there is an opportunity to explore alternative approaches for more efficient prediction and mitigation of PIDs in composites. This paper introduces a versatile framework for this purpose, which follows the subsequent workflow. First, the evolution of thermo-mechanical properties, including free strains and bending modulus, are characterized using a limited amount of Dynamic Mechanical Analysis (DMA) tests. Gaussian Process Regression (GPR), a probabilistic machine learning technique, is then employed to train models that predict these properties in reduced-order domains guided by integrated physics-based knowledge. The GPR models then serve as inputs in a low-fidelity finite element (FE) simulation scheme, based on 1D thermo-chemical and 2D closed-form thermo-mechanical analyses, to rapidly compute PIDs for composite parts in a defined design space. Subsequently, the low-fidelity virtual data is mapped to additional theory-guided domains and used to train GPR models. These GPR models are then iteratively retrained by incorporating high-fidelity 3D simulation data, based on 1D models and the Carrera Unified Formulation (CUF), and limited amounts of experimental data. In each retraining iteration, simulation datapoints are assigned uncertainties based on a Gaussian distance-decay weighing mechanism, creating an adaptive probabilistic model with a data-driven uncertainty structure. The method introduced in this work offers an alternative, cost-efficient, and broadly applicable framework for solving problems and potentially mitigating PIDs in composite parts.
Despite having a general understanding of PIDs, many manufacturers often struggle to accurately predict deformations in industrial settings. These difficulties primarily stem from limitations of traditional methods used for PID predictions, with one of the most notable constraints being the trade-off between fidelity/accuracy and time/cost. This trade-off frequently presents a dilemma for manufacturers when choosing a suitable prediction method, leading to attempts involving multiple methods and inadequate data. As a result, manufacturers are left with inaccurate predictions and seemingly uncontrollable PIDs.
Given challenges previously outlined, there is an opportunity to explore alternative approaches for more efficient prediction and mitigation of PIDs in composites. This paper introduces a versatile framework for this purpose, which follows the subsequent workflow. First, the evolution of thermo-mechanical properties, including free strains and bending modulus, are characterized using a limited amount of Dynamic Mechanical Analysis (DMA) tests. Gaussian Process Regression (GPR), a probabilistic machine learning technique, is then employed to train models that predict these properties in reduced-order domains guided by integrated physics-based knowledge. The GPR models then serve as inputs in a low-fidelity finite element (FE) simulation scheme, based on 1D thermo-chemical and 2D closed-form thermo-mechanical analyses, to rapidly compute PIDs for composite parts in a defined design space. Subsequently, the low-fidelity virtual data is mapped to additional theory-guided domains and used to train GPR models. These GPR models are then iteratively retrained by incorporating high-fidelity 3D simulation data, based on 1D models and the Carrera Unified Formulation (CUF), and limited amounts of experimental data. In each retraining iteration, simulation datapoints are assigned uncertainties based on a Gaussian distance-decay weighing mechanism, creating an adaptive probabilistic model with a data-driven uncertainty structure. The method introduced in this work offers an alternative, cost-efficient, and broadly applicable framework for solving problems and potentially mitigating PIDs in composite parts.