Fatigue property prediction and alloy design based on transfer learning for extremely small steel sample databases

Wei Xu*, Xiaolu Wei, Chenchong Wang, Chunguang Shen

The State Key Lab of Rolling and Automation, Northeastern University, Shenyang, 110819, China

EXTENDED ABSTRACT: Evaluation and prediction of fatigue properties for steels are of critical importance. However, traditional trial-and-error methods for fatigue strength oriented alloy design are costly and time consuming due to complicated fatigue tests, such as fatigue strength and S-N curves. Although data mining and machine learning have been widely employed for material design, replacing trial-and-error methods, adequate data accumulation for databases is required, while it is difficult for new material. Therefore, to reduce the costs of experimental validation and data accumulation, a transfer prediction framework for fatigue strength prediction and high-throughput alloy design was proposed based on the transfer learning concept and mechanics theory guidance. In the transfer prediction framework, machine learning models were first trained to estimate tensile properties. Then, based on the predicted tensile properties, transfer models were trained to estimate fatigue strength. The transfer prediction framework shows high accuracy for fatigue strength prediction with a high tolerance to the amount of data. By combining transfer prediction frameworks with evolutionary algorithms, a high-throughput fatigue strength oriented alloy design is achieved based on the use of only tens of fatigue data points as the training dataset. The transfer framework was further extended to the prediction of S-N curves. A transfer LSTM framework for S-N curve prediction was developed and the reversed torsion S-N curves prediction of low alloy steels were transferred from rotating bending S-N data. The universality of the framework at different amount of data and model parameters was further investigated. Additionally, it was also extended to the prediction of ultra-high cycle fatigue. This prediction framework could significantly reduce the cost of fatigue properties evaluation and realize the conversion between fatigue curves with different test costs. Based on the characteristic of adapting the transfer method to a small sample database, our transfer prediction framework is also portable for other systems. This research provides an example for the thorough integration of machine learning with physical metallurgy to solve the complex problem of property prediction and material design.