Session: 36-02 Surrogate Model-based applications

Paper Number: 153825

Physics Informed Machine Learning Model for Local Creep Prediction in Turbine Blade 

Gas turbine blades are an essential part of modern power generation systems, where they operate under high temperature, pressure and centrifugal load for prolonged period. To bear these extreme operating conditions, they are made up of nickel base superalloy. Over time, these operating conditions lead to creep deformation, which can adversely affect the part functionality. Creep is a time dependent phenomenon which causes plastic deformation of the component under prolonged exposure to high stress and temperature. Assessment of creep is hence an essential part of turbine blade design process. 

Conventionally, Creep assessment is carried out using Finite Element Methods which is a high-fidelity approach but it also time-consuming. This includes meshing of the component, applying boundary condition to model the physics of the problem, assigning material properties to the component, and solving the model based on differential equations that governs this phenomenon. Various material models namely strain hardening, time hardening, Norton Bailey Law, Garofalo Law etc. are used to describe the creep response of a given material, particularly metals and crystalline solids. These physics-based methods are quite accurate but take significant calculation time, generally 1-8 days, therefore limited calculation can be done.  

The motivation behind this work is to drastically reduce the calculation time for creep assessment which will augment the existing design process by exploring more design space, capturing uncertainties in boundary condition and further robustness studies. This becomes more imperative because of change in gas turbine operations driven by higher renewable penetration. 

This paper focuses on exploring physics informed machine learning techniques to predict the local creep strain in turbine blades. This provides flexibility in terms of analysing multiple models with different boundary conditions, material cases etc. To generate the dataset, creep analyses is performed on a turbine blade model under different temperature distribution condition. Multiple creep cases have been covered to generate the dataset for the model. The input features used for training this hybrid model is stress, strain, temperature across all the coordinates of the blade. The prepared dataset was further divided into training and testing dataset to train and later evaluate the performance and tune the hyperparameters of the model. Physics information is added in the neural network by incorporating the hard constraints and updating the loss function using creep material models. Along with physics information, ensemble neural network was studied as part of this work. The result obtained are promising which shows that physics informed machine learning can further augment the conventional physics-based approach for creep prediction.

Keywords: Hybrid Machine Learning model, Creep, Turbine blades, Local Temperature, Finite Element Analysis

Presenting Author: Vipin Pal Siemens Limited

Presenting Author Biography: I am Vipin Pal, senior executive engineer at Siemens Limited, India. I work as a Mechanical Integrity engineer in Gas Turbine domain. I have completed my Bachelor of Technology from UPES, Dehradun in Aerospace Engineering and later completed my master's from IIT Kanpur, India in Aerospace Engineering. I started my professional career working with Rolls-Royce Civil Aerospace Department as Trainee Engineer and currently I am working at Siemens Limited. My research interest includes design optimization of gas turbine using Machine Learning.

Authors:

Vipin Pal Siemens Limited
Amit Kumar Singh Siemens Limited
Souvik Chakraborty Indian Institute of Technology Delhi
Jason Abdallah Siemens Energy
Rishabh Shrivastava Siemens Limited