Aeropropulsive Design Optimization

Figure 1: The STARC-ABL concept uses boundary layer ingestion for improved aeropropulsive performance. A brief visualization of my work on this concept can be found at:

The development of future sustainable aircraft heavily relies on the design and integration of advancedpropulsion systems. My research focuses on enabling these technologies using aeropropulsive design optimization, in which the aerodynamic and propulsion system designs are optimized in a coupled manner.

For a NASA-funded project, I used coupled aeropropulsivemodels to study the benefit of boundary layer ingestion to the STARC-ABL concept (Fig. 1) [C2, C4, J6]. Currently, I am collaborating with NASA Glenn Research Center on creating a benchmark aeropropulsive model, which will be essentialfor the development of advanced modeling approaches in aeropropulsive design optimization [C6]. I also worked on the development of the multiphysics design optimization package MPhys [J8], which is built with NASA’s OpenMDAO framework. MPhys enables flexible formulation of MDO problems using high- and low-fidelity analysis tools. The extensible construction of MPhys will enable including additional disciplines such as structural and thermal analyses to the coupled aeropropulsive models I work with.

Nonlinear Solver Development

Robustness of the CFD solvers used in design optimization is extremely important. This is because we must perform hundreds of consecutive CFD simulations in a typical design optimization, and failures in the CFD solver cause the entire optimization framework to fail. These failures require human intervention, which increases the already high cost of these optimization problems even more.

To address this challenge, I developed the approximate Newton–Krylov (ANK) solver [J1J2] in the open-source CFD code ADflow [J4]. The ANK solver uses approximate residual routines in the context of a Jacobian-free Newton–Krylov solver algorithm. We further improve the robustness of the solver using pseudo-transient continuation. This solver provides a scalable and high-performance startup approach for RANS simulations where using a multigrid startup method is not available due to the complex mesh structures.

Besides its high performance, this solver is also extremely robust; we can even converge CFD simulations of aircraft at 90º angle of attack as shown in Figure 2, while many other solvers struggle with considerably less flow separation. The extreme robustness of this solver was essential for my work on aeropropulsive design optimization, and it also enabled the research of other Ph.D. students in our group. You can find more details about the ANK solver here.

Figure 2: The Common Research Model wing-fuselage configuration simulated at cruise conditions and 90º angle of attack. While this is not a physical solution, it illustrates the extreme robustness of the ANK solver [J2].

High Performance Computing

Figure 3: The speedup values using the blockette approach [C5] on the Stampede2 system.

Increased parallelism has been the main source of performance improvements in modern computers. In this sense, computers have been growing wider, not taller. However, our traditional approach of solving larger problems is by making our algorithms taller; expensive loops are stacked to build complex design optimization problems from standalone analyses. This approach will not scale to the levels of computational performance required to solve design problems that are increasingly more expensive. Furthermore, the new processor architectures heavily rely on heterogeneous and multi-level memory hierarchies, which makes achieving optimal performance even more difficult.

To address some of these challenges, I worked on the modernization of ADflow for modern high-performance computing resources by efficiently using memory hierarchies. To fully utilize the available memory bandwidth on the nodes of the Stampede2 supercomputer, we implemented a cache-blocking method and achieved up to 3 times speedup in the core routines of ADflow [C5]. I have also worked on a matrix-free linear preconditioner in ADflow. This approach reduces the total memory usage and increases the operational intensity of the linear solvers used.

Geometry Parameterization in MDO

The challenges in computer-aided design (CAD) based geometry parameterization with CFD-based design optimization has been a barrier for wide adoption of numerical design optimization in industry. The traditional “free form deformation” approach that is widely used in optimization falls short when we want to directly work with CAD models of the configurations.

To address this challenge, I worked on the development of high-performance wrappers for the OpenVSP and ESP tools to integrate them in the MACH-Aero framework. My approach enabled a fully memory-based wrapping of these tools to be used in design optimization. Furthermore, I implemented a parallel finite-differencing approach to obtain design sensitivities from these geometry tools where analytic sensitivities are not always available.

Finally, a significant challenge in using CAD-based geometry parameterization is due to how the component intersections are handled as the design changes during optimization. To address this challenge, I developed a new method that deforms the CFD mesh near intersecting aircraft components as the design changes (Fig. 4) [J5C3]. This method enables the use of CAD-based geometric parameterization in design optimization. Besides this, I supported the development of a method to formulate arbitrary spatial packaging constraints [J3], which will be critical in addressing the challenges of integrating energy storage systems in the airframe, such as batteries and hydrogen tanks.

Figure 4: The surface mesh deformation algorithm presented in [J5] is used to adjust the mounting angle of the wing with respect to the fuselage of the DLR-F6 configuration.