A machine-learning-based model for predicting crack arrest fracture toughness through literature data mining

Machine learning (ML) is transforming fracture mechanics research by offering unprecedented capabilities for analyzing high-throughput, multi-modal, and multi-fidelity data. Recent work has generated large amounts of new fracture data through experiments and simulations and extracted useful insights from them. However, the vast body of fracture data accumulated over the past century remains largely untapped, primarily because it is scattered throughout the literature without systematic organization. Our central hypothesis is that these historical data contain valuable insights that can be unlocked using ML. To test this hypothesis, we present a case study on crack arrest fracture toughness, a critical yet poorly understood property in fracture mechanics. We compiled a comprehensive dataset of crack arrest toughness for a wide range of structural steels through a thorough literature review and analyzed it using neural network (NN)-based methods, including feature-importance assessment and high-dimensional regression. Our analysis shows that elements such as carbon and manganese exert a stronger influence on crack arrest toughness than others such as copper, and that temperature also plays a critical role. We further developed an NN-based model capable of predicting crack arrest toughness from these factors with an error of 11.8%. This study demonstrates the substantial opportunities for advancing fracture mechanics by mining the vast body of historical literature data, while also highlighting the challenges associated with their fragmented and multi-fidelity nature.