Mavris, Dimitri N.

Associated Organization(s)
ArchiveSpace Name Record

Publication Search Results

Now showing 1 - 3 of 3
  • Item
    Shock Wave Prediction in Transonic Flow Fields using Domain-Informed Probabilistic Deep Learning
    (Georgia Institute of Technology, 2024-01) Mufti, Bilal ; Bhaduri, Anindya ; Ghosh, Sayan ; Wang, Liping ; Mavris, Dimitri N.
    Transonic flow fields are marked by shock waves of varying strength and location and are crucial for the aerodynamic design and optimization of high-speed transport aircraft. While deep learning methods offer the potential for predicting these fields, their deterministic outputs often lack predictive uncertainty. Moreover, their accuracy, especially near critical shock regions, needs better quantification. In this paper, we introduce a domain-informed probabilistic (DIP) deep learning framework tailored for predicting transonic flow fields with shock waves called DIP-ShockNet. This methodology utilizes Monte Carlo Dropout (MCD) to estimate predictive uncertainty and enhances flow field predictions near the wall region by employing the inverse wall distance function (IWDF) based input representation of the aerodynamic flow field. The obtained results are benchmarked against the signed distance function (SDF) and the geometric mask input representations. The proposed framework further improves prediction accuracy in shock wave areas using a domain-informed loss function. To quantify the accuracy of our shock wave predictions, we developed metrics to assess errors in shock wave strength and location, achieving errors of 6.4% and 1%, respectively. Assessing the generalizability of our method, we tested it on different training sample sizes and compared it against the proper orthogonal decomposition (POD)-based reduced order model (ROM). Our results indicate that DIP-ShockNet outperforms POD-ROM by 60% in predicting the complete transonic flow field.
  • Item
    Design Space Reduction using Multi-Fidelity Model-Based Active Subspaces
    (Georgia Institute of Technology, 2023-06) Mufti, Bilal ; Perron, Christian ; Gautier, Raphaël ; Mavris, Dimitri N.
    The parameterization of aerodynamic design shapes often results in high-dimensional design spaces, creating challenges when constructing surrogate models for aerodynamic coefficients. Active subspaces offer an effective way to reduce the dimensionality of such spaces, but existing approaches often require a substantial number of gradient evaluations, making them computationally expensive. We propose a multi-fidelity, model-based approach to finding an active subspace that relies solely on direct function evaluations. By using both high- and low-fidelity samples, we develop a model-based approximation of the projection matrix of the active subspace. We evaluate the proposed method by assessing its active subspace recovery characteristics and resulting model prediction accuracy for airfoil and wing drag prediction problems. Our results show that the proposed method successfully recovers the active subspace with an acceptable model prediction error. Furthermore, a cost vs. accuracy comparison with the multi-fidelity gradient-based active subspace method demonstrates that our approach offers comparable predictive performance with lower computational costs. Our findings provide strong evidence supporting the usage of the proposed method to reduce the dimensionality of design spaces when gradient samples are unavailable or expensive to obtain.
  • Item
    A Multi-Fidelity Approximation of the Active Subspace Method for Surrogate Models with High-Dimensional Inputs
    (Georgia Institute of Technology, 2022-06) Mufti, Bilal ; Chen, Mengzhen ; Perron, Christian ; Mavris, Dimitri N.
    Modern design problems routinely involve high-dimensional inputs and the active subspace has been recognized as a potential solution to this issue. However, the computational cost for collecting training data with high-fidelity simulations can be prohibitively expensive. This paper presents a multi-fidelity strategy where low-fidelity simulations are leveraged to extract an approximation of the high-fidelity active subspace. Both gradient-based and gradient-free active subspace methods are incorporated with the proposed multi-fidelity strategy and are compared with the equivalent single-fidelity method. To demonstrate the effectiveness of our proposed multi-fidelity strategy, the aerodynamic analysis of an airfoil and a wing are used to define two application problems. The effectiveness of the current approach is evaluated based on its prediction accuracy and training cost improvement. Results show that using a low-fidelity analysis to approximate the active subspace of high-fidelity data is a viable solution and can provide substantial computational savings, yet this is counterbalanced with slightly worse prediction accuracy.