Our group develops computational techniques for aerospace decision-making. We are focused on the novel methodological development behind decision-making methods and, thus, work on a diverse range of applications, with the majority of it being aerospace systems. As shown in the figure below, our research hinges on three main focal points: approximation theory, uncertainty quantification, and numerical optimization. This page provides details on currently active research projects.
Project 1: Gaussian processes, Bayesian optimization, and friends.
Anchored on (deep) Gaussian processes (GP) models and Bayesian decision theory, we actively develop adaptive methods (sometimes known as “active learning”) to solve variety of problems involving expensive oracles, which include (i) constrained single objective optimization, (ii) constrained multiobjective optimization, (iii) high-dimensional surrogate modeling, (iv) contour finding and failure probability estimation, and (v) quadrature for uncertainty quantification. We are primarily focused on three main goals within this project:
- faster empirical and theoretical convergence rates,
- efficient “batch” acquisitions for asynchronous parallel evaluation, and
- hardware acceleration via distributed GPU computing.
Recent work
- qPOTS: Our recent (and ongoing) work in this area is qPOTS, an efficient multiobjective Bayesian optimization method based on Pareto optimal Thompson sampling. qPOTS demonstrates unprecedented computational efficiency and accuracy compared to several methods in the state of the art (see graphic below and find out more here).
qPOTS is a fast and accurate multiobjective Bayesian optimization method that is based on Thompson sampling.
Relevant publications:
- Renganathan, S. A. (2023). qPOTS: Efficient batch multiobjective Bayesian optimization via Pareto optimal Thompson sampling. arXiv preprint arXiv:2310.15788.
- Booth, A. S., Renganathan, S. A., & Gramacy, R. B. (2023). Contour Location for Reliability in Airfoil Simulation Experiments using Deep Gaussian Processes. arXiv preprint arXiv:2308.04420.
- Renganathan, S. A., Rao, V., & Navon, I. M. (2023). CAMERA: A method for cost-aware, adaptive, multifidelity, efficient reliability analysis. Journal of Computational Physics, 472, 111698. https://doi.org/10.2514/6.2022-0390
- Renganathan, A., Rao, V., & Navon, I. (2022). Multifidelity Gaussian processes for failure boundary and probability estimation. In AIAA SCITECH 2022 Forum (p. 0390). https://doi.org/10.2514/6.2022-0390
- Renganathan, S. A., Larson, J., & Wild, S. M. (2021). Lookahead acquisition functions for finite-horizon time-dependent bayesian optimization and application to quantum optimal control. arXiv preprint arXiv:2105.09824. https://arxiv.org/abs/2105.09824
Project 2: Multifidelity surrogate modeling, optimization, and uncertainty quantification
Modern decision-making for design depends on computer models of the physical system under consideration. Often, there is no one model of the system that we can use, but a set of models that trade accuracy for computational cost, as shown in the figure below. Building on project 1, we solve several decision-making problems which can exploit lower fidelity “cheaper” models to gain computational efficiency without any significant loss of accuracy. Specifically, we are interested in solving problems in a cost-aware manner such that adaptive decisions are made while accounting for the computational cost of querying a model at a specific fidelity.
Recent work
CAMERA: Recently, we came up with a method, CAMERA: A method for Cost-Aware Multifidelity Efficient Reliability Analaysis (read more about it here), where the estimation of failure probabilities (a.k.a., reliability analysis) can be performed a lot more computationally cheaply with multifidelity models. The figures below shows the demonstration of CAMERA on a turbine blade thermal stress problem.
Relevant publications:
- Renganathan, S. A., Rao, V., & Navon, I. M. (2023). CAMERA: A method for cost-aware, adaptive, multifidelity, efficient reliability analysis. Journal of Computational Physics, 472, 111698. https://doi.org/10.2514/6.2022-0390
- Renganathan, A., Rao, V., & Navon, I. (2022). Multifidelity Gaussian processes for failure boundary and probability estimation. In AIAA SCITECH 2022 Forum (p. 0390). https://doi.org/10.2514/6.2022-0390
Project 3: Machine learning for reduced order modeling in aerodynamics
We are actively involved in learning reduced order models of nonlinear aerodynamic flows, particularly for aircraft configurations and wind energy applications. Sometimes, we combine this with project 1, to enable active learning for surrogate modeling.
Relevant publications:
- Rajaram, D., Puranik, T. G., Renganathan, A., Sung, W. J., Pinon-Fischer, O. J., Mavris, D. N., & Ramamurthy, A. (2020). Deep Gaussian process enabled surrogate models for aerodynamic flows. In AIAA Scitech 2020 Forum (p. 1640).
- Renganathan, S. A., Liu, Y., & Mavris, D. N. (2018). Koopman-based approach to nonintrusive projection-based reduced-order modeling with black-box high-fidelity models. AIAA Journal, 56(10), 4087-4111.
- Renganathan, S. A. (2020). Koopman-based approach to nonintrusive reduced order modeling: Application to aerodynamic shape optimization and uncertainty propagation. AIAA Journal, 58(5), 2221-2235.
- Renganathan, S. A., Maulik, R., & Rao, V. (2020). Machine learning for nonintrusive model order reduction of the parametric inviscid transonic flow past an airfoil. Physics of Fluids, 32(4), 047110.
- Ashwin Renganathan, S., Maulik, R., Letizia, S., & Iungo, G. V. (2022). Data-driven wind turbine wake modeling via probabilistic machine learning. Neural Computing and Applications, 1-16.
Research sponsors
Our research has so far been funded by the U.S. Department of Energy (Lab Directed Research and Development program), Penn State College of Engineering, and the Penn State Institute of Computational and Data Sciences (ICDS). All our computations are carried out on Penn State’s in-house supercomputer Roar Collab on 48-core, 256GB CPU nodes, and NVIDIA A100 GPUs.