TY - GEN
T1 - Physics-guided machine learning for multifactor fatigue analysis and uncertainty quantification
AU - Chen, Jie
AU - Liu, Yongming
N1 - Funding Information:
The work in this study is partially supported by fund from NSF (award No: 1536994). The support is greatly acknowledged.
Publisher Copyright:
© 2021, American Institute of Aeronautics and Astronautics Inc, AIAA. All rights reserved.
PY - 2021
Y1 - 2021
N2 - The Probabilistic Physics-guided Neural Network (PPgNN) model is proposed for multifactor fatigue data analysis with uncertainty quantification. The proposed PPgNN avoids the limitations of the commonly used physics-based regression models and the classical neural network methods. Compared with regression models, it is more convenient to incorporate the effect of fatigue influencing factors other than stress level, which benefits the multi-factor fatigue analysis. Improvements can be achieved to avoid overfitting and extrapolation issue commonly occurring in the classical neural network by guiding the machine learning process with known physics knowledge. This is accomplished by a constrained optimization process and proper neural network architecture design. With the proposed machine learning method, both the mean and variance variation can be estimated from the data, which enables uncertainty quantification. The PPgNN is validated using the experimental data. Probabilistic S-N curved are produced using the proposed method. To illustrate the impact of the physics guidance on the machine learning process, the results from the classical neural network without physics guidance and PPgNN are compared. The proposed Probabilistic Physics-guided Neural Network is shown to generate both accurate and physically consistent results. By reconstruction the network architecture, this method is not restricted to fatigue data and has the potential for the analysis of other types of data.
AB - The Probabilistic Physics-guided Neural Network (PPgNN) model is proposed for multifactor fatigue data analysis with uncertainty quantification. The proposed PPgNN avoids the limitations of the commonly used physics-based regression models and the classical neural network methods. Compared with regression models, it is more convenient to incorporate the effect of fatigue influencing factors other than stress level, which benefits the multi-factor fatigue analysis. Improvements can be achieved to avoid overfitting and extrapolation issue commonly occurring in the classical neural network by guiding the machine learning process with known physics knowledge. This is accomplished by a constrained optimization process and proper neural network architecture design. With the proposed machine learning method, both the mean and variance variation can be estimated from the data, which enables uncertainty quantification. The PPgNN is validated using the experimental data. Probabilistic S-N curved are produced using the proposed method. To illustrate the impact of the physics guidance on the machine learning process, the results from the classical neural network without physics guidance and PPgNN are compared. The proposed Probabilistic Physics-guided Neural Network is shown to generate both accurate and physically consistent results. By reconstruction the network architecture, this method is not restricted to fatigue data and has the potential for the analysis of other types of data.
UR - http://www.scopus.com/inward/record.url?scp=85100314110&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85100314110&partnerID=8YFLogxK
M3 - Conference contribution
AN - SCOPUS:85100314110
SN - 9781624106095
T3 - AIAA Scitech 2021 Forum
SP - 1
EP - 11
BT - AIAA Scitech 2021 Forum
PB - American Institute of Aeronautics and Astronautics Inc, AIAA
T2 - AIAA Science and Technology Forum and Exposition, AIAA SciTech Forum 2021
Y2 - 11 January 2021 through 15 January 2021
ER -