Artificial Intelligence-Powered Probabilistic Damage Tolerance Approach for Impeller Mechanical Assessment

Abstract This energy transition has the main target to move from a fossil fuel-based energy market to a cleaner and more versatile one, making use of renewable sources. It involves not only energy production itself, but also the way it is distributed, stored, and managed. Centrifugal compressors used to fulfill energy transition encounter a specific technical challenge: renewable energy-driven services are characterized by a significant intermittency of operation, from a daily start/stop cycling to a periodic or seasonal operation. This paper presents a probabilistic damage tolerance approach for centrifugal compressor impellers, addressing fatigue strength requirements under these challenging conditions. A Monte Carlo simulation workflow is utilized to predict fatigue crack growth rates based on material properties, defect sizes, and rotor speed history. Despite significant algorithm optimization, the computation time remains high, limiting its practical application. To overcome this, a hybrid approach employing machine learning is proposed. A classifier model predicts failure likelihood, while a regression model estimates crack propagation for low-risk samples. This method reduces the number of simulations by 50%, with comparable accuracy to random sampling, and significant reductions in computing time. Validation against various rotor designs confirms minimal variation in predicted results, demonstrating the efficacy and efficiency of this approach.