Research Projects

An optimal design of turbine blades is crucial for the performance and efficiency of gas turbine engines. Even a modest 1% increase in the efficiency of low-pressure turbine (LPT) blades can lead to a reduction in direct operating cost (DOC) by approximately 0.26%. For a country like India, with around 800 commercial flight operations per day, this translates to annual savings on the order of 20 million USD. However, improving the aerodynamic design of LPT blades remains a significant challenge due to the complexity of the flow field around them. These blades are typically highly curved to maximize energy extraction and operate under harsh conditions, including extreme temperatures, pressures, and high levels of turbulence, which vary dynamically across flight phases such as takeoff, cruise, and landing. As a result, the flow over turbine blades can exhibit a wide range of phenomena—including boundary layer separation, laminar-turbulent transition, and relaminarization—earning it the description of a true “fluid-dynamical zoo”. Capturing these flow features accurately is essential for informed design, but remains extremely challenging using conventional simulation or experimental approaches alone. Experimental measurements are considered the most direct and reliable source of information for these flows. However, they are typically sparse in both space and time, limited by sensor resolution and accessibility. Quantities such as wall shear stress and heat-transfer coefficients are difficult to measure directly, and resolving near-wall regions or flow separation accurately remains challenging due to instrumental limitations and data coarseness. In contrast, CFD simulations provide detailed, high-resolution predictions across the full domain and enable access to quantities that are often inaccessible experimentally. However, these predictions depend heavily on turbulence models and boundary conditions, which introduce uncertainty. In this context, Physics-Informed Machine Learning (PIML)—offers a promising alternative. By incorporating physical laws, such as partial differential equations (PDEs) and boundary conditions, directly into the learning process, PIML enables the fusion of sparse experimental measurements with computational physics models. This hybrid approach allows for the reconstruction of high-resolution, physically consistent flow fields, even in regions where measurements are limited or simulations are unreliable. We focus on Physics-Informed Neural Networks (PINNs) that embed physical laws in the form of differential equations into the learning process, enabling both forward and inverse problem-solving. While PINNs have shown success in various domains, their application to fluid mechanics—especially in turbulent, nonlinear, and chaotic regimes—remains limited. Existing studies have primarily demonstrated PINN effectiveness in low Reynolds number laminar flows, where the governing physics is simpler and the solution space is more regular. However, turbine blade flows present a far more complex flowfield, as described earlier, where PINNs struggle with accuracy even to predict qualitative features. This proposal aims to address this gap by developing a hybrid framework that integrates sparse experimental data and physics-based knowledge from CFD models within the PINN architecture, to enable fast and reliable prediction of full flowfields. The approach holds the potential to significantly accelerate India’s gas turbine design program, such as those by DRDO, by enabling faster and more cost-effective aerodynamic design and optimization cycles of blades.