Session: 37-01 Neural Network based surrogate models and optimization

Paper Number: 100746

100746 - Surrogate Models for 3d Finite Element Creep Analysis Acceleration 

A gas turbine and especially its blades is a heart of any gas power station and thus it is critical to maximize their long design life as a key component for successful utilization. Blades weaknesses could affect the whole process of the energy generation, contractual penalty shall be paid in this case and company reputation can be affected. Therefore, it is important to find out the weakness of blades as quickly as possible. One of the central causes of weakness is creep, that is the persistent deformation of material under the influence of mechanical stresses and temperature.

 

Experience gained from real field issues show that high fidelity creep models must be used since the early stages of component design and by field incident investigations. One of such analysis tools, the Theta Projection Model (TPM) delivers accurate but computationally expensive results – depending on the mesh quality and fidelity of the geometry. The method requires 2-8 days of computing time to provide only a single solution about the mechanical behavior and capacity of just one blade geometry.

 

To accelerate computation, in this paper we investigate the integration of machine learning methods in design and process chains.  Specifically, we create surrogate models, based on neural networks, that predict, given a point cloud that represents the blade topology, the creep impact for the full topology at every time step. Additionally, the surrogate will receive contextual variables as input, such as the metal temperature and geometry parameter.  Training a surrogate on such data enables a transfer to new temperature and geometry settings.  We evaluate the proposed approach on data from turbine blade analysis and evaluate the surrogate model to a set of baselines. The data is generated using automated tool chains. Our main finding is that a surrogate model provides accurate predictions of the creep impact in many cases.

 

The results indicate that integrating machine learning in the design process has the potential to decrease the computing time from days down to minutes. Another important benefit of such synergy is very fast 3D creep predictions which increase responsiveness to potential utilization issues. This is required, for example, for Creep Capability, TMF & HCF. Furthermore, the acceleration of 3D creep simulations assessment would enable the integration of non-linear material behavior into Multi-Disciplinary Optimization (MDO) applications.  Therefore, the creep simulations would provide not only closer to the real issues but also more efficient solutions.

Presenting Author: Jason Abdallah Siemens Energy

Presenting Author Biography: Experienced Senior Mechanical Design Engineer with a demonstrated history of working in the professional large gas turbines industry. Strong business development professional with a Diploma (2012) focused in Aerospace Engineering from Technical University of Berlin.

Authors:

Jason Abdallah Siemens Energy
Stefan Depeweg Siemens AG
Maria Kuznetsova Siemens Energy
Behnam Nouri Siemens Energy