Physics-based machine learning for computational fracture mechanics

Accurately predicting crack initiation, growth, and long-term fatigue degradation remains one of the central challenges in computational mechanics. Traditional physics-based models, such as phase-field and continuum damage formulations, provide detailed and physically sound descriptions of brittle and ductile fracture. However, these methods become prohibitively expensive when applied to large-scale structures, long-term cyclic loading, or high-cycle fatigue. This makes the development of efficient, physically consistent data-driven methods essential for ensuring the safety and sustainability of engineering systems such as aircraft, bridges, offshore structures, and wind turbines.

In this project, we plan to develop a physics-based machine learning framework that embeds core physical principles directly into the neural network architecture. Using a feedforward neural network designed to satisfy governing equations and thermodynamic constraints, the framework will integrate mechanics, constitutive behavior, and energy balance at the architectural level. Synthetic data generated from finite element–based phase-field fracture simulations will be used to train the model, initially focusing on homogeneous one-dimensional responses for both brittle and ductile materials. Special emphasis will be placed on learning the evolution of elastic energy, dissipated work, and fracture characteristics in a way that preserves physical interpretability and consistency, addressing key limitations of classical ML approaches that rely purely on data and lack embedded physical guarantees.

Building on this foundation, the framework will be extended to handle cyclic and fatigue loading, enabling the prediction of fatigue damage accumulation, crack nucleation, and fatigue-driven crack growth. This will open a pathway toward efficient, physics-informed surrogate models capable of capturing long-term degradation without the computational overhead of full finite element simulations. The ultimate goal is to perform high-cycle and very-high-cycle fatigue simulations in both 2D and 3D settings at a fraction of the traditional computational cost, while retaining the essential physics of brittle and ductile fracture.

Through this physics-integrated design, the project aims to establish a new class of ML-based fracture models that remain reliable under limited data, generalize across loading conditions, and provide a physically grounded alternative for fatigue life prediction in modern engineering applications.