Abstract
Gas turbines blades are subjected to significantly high temperature, pressure, and centrifugal load for prolong period. To withstand this severe operating condition, turbines blades are made up of nickel base superalloy. However, still we observe that creep is one of the dominant failure modes for turbine blade and could cause significant damage to the turbine section. Creep is the progressive time-dependent in elastic deformation resulting from high stress and temperature. Because of above reasons, creep evaluation becomes an imperative part of the turbine design. Creep analysis involves multiple discipline such as aerodynamic, heat-transfer and mechanical integrity. Additionally, conventional creep analysis using finite element model requires defeaturing of models, meshing of geometry, boundary condition application, material definition, solving the model and post processing. The process involves a numerical solution of differential equations that governed these physical phenomena. The above steps need significant time, computational resource, and manpower. The mentioned limitation could be overcome by leveraging data driven techniques for calculating maximum creep strain using system level parameters. In this work, machine learning techniques are explored to predict maximum creep strain in gas turbine blades. Machine learning models will use FEA results as input to predict the maximum creep strain values. Since creep strain is a continuous variable, hence regression-based ML model will be studied. The input parameter used for building the machine learning models are coolant temperature, coolant mass flow rate, hot side mass flow rate, turbine inlet temperature along with blade material. Among all the data, 75% data were used for training purpose and 25% data were used for evaluating the machine learning model performance. With the developed model, time required for creep prediction is reduced significantly. This provides greater flexibility to the designer in terms of evaluating multiple boundary conditions, materials option, coating, and cooling optimization. A comparative study is presented between linear & polynomial regression, random forest, and support vector regressor models. To reduce the prediction error, hyperparameter optimization is carried out for all the studied models. Additionally, a feature importance study is performed to assess the impact of each parameter on the creep strain prediction. Findings from the study show that linear regression model generates the least root mean squared error. Study also shows that Turbine inlet temperature has the maximum impact on creep strain, which also aligns with our physics-based understanding of creep phenomenon.