Nowadays, cross-linked elastomers play a significant role in several industries such as aerospace, construction, transportation, marine, aeronautics, and automotive due to excellent flexibility, toughness, form-ability,and versatility. During their intended service-life, the material is supposed to sustain aggressive environ- mental damages induced by water infusion, temperature, and solar ultraviolet radiation (UV) during their operation, which affects their durability and properties. A reliable design of rubber components to prevent early failure by environmental degradation requires digital simulations by means of high-fidelity thermo-mechanical constitutive models that can simulate the adverse effects of aging on mechanical, electrical, thermal, and failure properties of polymers. So far, most aging models are developed by coupling hyperelastic constitutive models with single-kinetic degradation models, to demonstrate the decay of materials during aging. However, a more detailed modeling approach can be achieved through modular continuum-based damage models that integrate the finite strain theory and thermo-mechanical degradation models. Rubber elasticity theory is driven partly based on (i) statistical mechanics at micro-scale (ii) Phenomenological Modeling at Meso-scale for modeling of the network (iii) Continuum Mechanics at Macro-scale to model the material. So, hyperelastic models fall into three main categories: the phenomenological approach, the micro-mechanical approach, and the data-driven approach. Recently, the emergence of machine-learned (ML) models has attracted much attention. The first generation of "black-box" ML models as another type of phenomenological model was proposed to model the mechanical behavior of rubbery media. In solid mechanics, stress–strain tensors are only partially observable in lower dimensions. Thus, obtaining data to feed a black-box ML model is exceptionally challenging. Thus, these approaches soon become obsolete due to the high demand for data for training, and the lack of constraint on their output margins. The issue can be resolved in a new generation of ML models which is inspired by physics-informed neural networks (PINN) which infuse physics-based knowledge into the black-box models. Here, we modify PINN models to develop hybrid frameworks that can address the limitations of both phenomenological and micro-mechanical models by obtaining micro-structural behavior from the macroscopic experimental data set. The objective of this dissertation is to provide a new approach for reduced-order physics-based Data- driven modeling of multi-stressor damage in elastomers by infusing Knowledge into a neural network. Thefollowing are the major thrusts of our research in the proposed dissertation:• To design a systematic approach to reduce order of the constitutive mapping and address the datavolume problem for training.• To incorporate background knowledge from polymer physics, continuum mechanics, and thermodynamics into the neural networks and constraint the solution space.• To develop a neural network to predict various inelastic effects which is far less data-dependent, moreinterpretable than current PINN, and uses a knowledge-confined solution space.• To validate our proposed hybrid framework based on limited data to describe the relationship betweenelastomeric network mechanics and environmental degradation.To go into further detail, the model has been successfully developed and validated in five different dam-age scenarios which describe the evolutionary process of developing the final platform. These steps are as follows, (I) providing a model for polymers in non-extreme environments to capture the dependence of elastomer behavior on loading conditions such as strain rate and temperature, as well as compound morphology factors such as filler percentage and crosslink density, (II) developing a model for single mechanism aging, i.e. thermal aging, or hydrolytic aging, (III) developing a model to capture accumulation damages of fatigue and thermo-aging, (IV) introducing Physics informed neural networks (PINNs) to simulate multiple stiff, and semi-stiff ODEs that govern Pyrolysis and Ablation, and (V) developing a Bayesian surrogate constitutive model to estimate failure probability of elastomers. The models used in the proposed platform are the first hybrid models developed and validated for polymer components and thus, bring great novelty and value to the industry. The model proposed in this work can significantly improve the design process of polymeric components by predicting the reliability, durability, and performance loss of materials based on the projected mechanical and environmental loading conditions. Such knowledge can significantly reduce the design cost, reduce the number of reliability tests needed, reduce the maintenance costs and overhauls, and most importantly prevent unexpected catastrophic failures.
Read
