In 2021, a United Airlines aircraft travelling from Denver to Honolulu experienced engine failure, culminating in a fire. Following an investigation by the National Transportation Safety Board, it was determined that a fractured fan blade in the engine had sustained damage consistent with metal fatigue, causing the engine failure.
To better predict—and ultimately avoid—incidents of fatigue failure, team of engineers from Johns Hopkins University’s Department of Civil and Systems Engineering and Pratt & Whitney, an aerospace manufacturer, has developed a sophisticated multiscale model integrating physics-based modeling, machine learning and uncertainty quantification to predict fatigue nucleation in metallic parts before the appearance of cracks detected by conventional non-destructive evaluation methods.
Fatigue failure in metallic parts, the weaking and eventual breaking of the material, is one of the costliest engineering challenges in the aerospace, automotive, transportation, construction, and industrial sectors, with an estimated expense of more than 4% of the total gross domestic product of developed countries, like the U.S. and U.K.
“Fatigue is a stealthy problem,” says principal investigator, Somnath Ghosh. “You typically don’t know what what’s going on beyond the visible surface area, and you don’t know when or where fatigue cracks will appear, which could have disastrous consequences, especially when human lives are at stake.”
Published in Nature Communications, the study presents a prognosis tool for fatigue, termed parametrically upscaled constitutive and crack nucleation models, or PUCM-PUCNM for short, that directly links extreme microstructural events with macroscopic failure and overcomes the many challenges encountered through conventional methods of fatigue prediction. By integrating physics-based modeling, machine learning, temporal acceleration and probabilistic analysis, PUCM-PUCNM can predict changes that happen over time due to repetitive use and prolonged loading, as well as the location of fatigue failure before it happens.
The team’s study presents their successful use of the model for fatigue prediction in titanium alloys which are frequently used in aircraft engines.
The team began by analyzing titanium alloys to understand the polycrystalline microstructure. They then created a 3D model of the microscopic structure and used physics-based models to simulate how the metal would deform when exposed to repeated or extended stresses. Because fatigue can take thousands or millions of cycles of loading before failure, the researchers applied temporal acceleration to speed simulations by more than three orders of magnitude to evaluate a greater number of simulations. Machine learning was applied to generate macroscopic material response models, mapping microscopic behavior over time to the moment of fatigue failure. Lastly, due to the variance in the microstructures of different titanium alloys, the model incorporates the probability of how likely a crack is to form after a certain number of cycles of loading.
“This approach is the outcome of nearly 15 years of research supported by federal funding and tested in collaboration with industry partners like Pratt & Whitney, GE Aerospace and Rolls-Royce,” Ghosh says. “These companies are seeing promising results and are considering the model in their engineering design. They’re also sharing non-proprietary data to help us continue to advance and validate our models.”
He says that the models are suitable for relatively easy industrial use, which makes them particularly attractive.
To determine the accuracy of PUCM-PUCNM’s predictions, the team completed three case studies, and in each instance compared the model’s predictions with observations collected from previous fatigue failure studies. The first two case studies showed that PUCM-PUCNM could accurately predict dwell fatigue of a notched metal specimen and cyclic loading of a compressor disk. The third case study demonstrated that when calibrated on small-scale sections, PUCM-PUCNM could be used to predict behavior in a full-size part with more accuracy than existing methods.
Enhanced fatigue predictions from the new modeling tool have both economic and safety implications. With improved prognosis technologies, engineers can design components to extend their lifespan, optimize maintenance schedules, and avoid unplanned failures.
“Improved product life and performance could amount to a cost savings of several hundred million dollars, globally, through federal government and industry use,” Ghosh says.
While the model is applicable across materials and manufacturing methods, including additively manufactured alloys which are becoming more common in the aerospace and defense industries, the team is also exploring the use of their predictive model in damage sensing and non-destructive evaluation by integrating it with additional AI tools, such as neural operators.
“We’re advancing the frontier in this area, and the results could be significant,” says Ghosh. “Industry participation is telling that this technology is directly translatable to users for substantial gains. Our study is a demo case, but it doesn’t end there. We’ve shown that this is a meaningful way of coupling AI with physics-based models and that holds great promise for future investments at federal and industry levels.”
This work was supported by the Air Force Office of Scientific Research’s Structural Mechanics and Prognosis Program and the Air Force’s Metals Affordability Initiative and SPARTA program. Additional study participants include Johns Hopkins University’s Kishore Appunhi Nair, Tawqeer Nasir Tak and Shravan Kotha, and Pratt & Whitney’s Adam Pilchak, Vasisht Venkatesh and David Furrer.