Session: 36-07 Machine Learning & Artificial Intelligence Methods - Part 2

Submission Number: 177536

Surrogate Model Based Optimization of Surface Morphed Gyroid Lattices in Internal Cooling Channels 

For gas turbine power cycles, increasing the turbine inlet temperature (TIT) improves the overall cycle efficiency. This challenges the material limitations of turbine blades and risks thermal failure. Internal cooling channels are critical to prevent material failure of such blades. Turbulators such as ribbed channels in mid-chord regions and pin fin arrays in trailing edges are typically utilized to improve cooling. It is critical to optimize the performance of such cooling channels, to reduce pumping power loss associated with diverted compressor air used for internal cooling, while maximizing thermal performance. The current study looks at gyroid triply periodic minimal surface (TPMS) lattices in place of traditional turbulators. Although the baseline gyroid equations and other lattices have been studied previously, a parameterized surface optimization using underlying level set equations has not been attempted as much. By parameterizing the gyroid equations using multipliers, morphed lattices are formed with varying degrees of asymmetry. An asymmetric shape created with the aim of end-wall heat removal can potentially provide higher heat transfer enhancement compared to a symmetric gyroid lattice in all directions. A Gaussian process (GP) machine learning surrogate, with Bayesian optimization, is then used to optimize the shape of the baseline gyroid TPMS, with the aim of maximizing thermal efficiency. This parameter combines Nusselt number and pressure drop to evaluate cooling channel effectiveness. Upto 20% improvement in thermal efficiency was observed for a modified shape, compared to the symmetric baseline gyroid design. A GP surrogate was trained on an initial set of samples, for which a high-fidelity RANS-CFD simulation was run per design data point. This was followed by a Bayesian optimization stage, which suggests new design points iteratively, based on an acquisition function, till the computation time ran out. The exploration vs exploitation approach of Bayesian Optimization ensured data efficiency and low computational cost, while searching for an optimum design with the highest thermal efficiency. 

 

Presenting Author: Shinjan Ghosh University of Central Florida

Presenting Author Biography: Shinjan is currently a postdoctoral scholar at the University of Central Florida (Orlando,), Center For Advanced Turbomachinery and Energy Research (CATER). His areas of focus are generative design, surrogate model based design and Physics informed AI with applications to thermal-fluid CFD. In the past, he has worked as a research scientist at Siemens Research (Princeton, NJ), where he focused on graph neural networks and PINNs for turbulent flows, and an intern at Siemens DI software (Orlando), where he worked on Multiphysics Topology Optimization.

Authors:

Shinjan Ghosh University of Central Florida
Abhilash Prasad University of Central Florida
Marcel Otto University of Central Florida
Erik Fernandez University of Central Florida
Jayanta Kapat University of Central Florida