Browsing by Author "Yukselen, Mahmut A."

Now showing 1 - 2 of 2
  • Thumbnail Image
    ItemOpen Access
    Design of a deep learning based nonlinear aerodynamic surrogate model for UAVs
    (AIAA, 2020-01-05) Karali, Hasan; Demirezen, Mustafa U.; Yukselen, Mahmut A.; Inalhan, Gokhan
    In this paper, we present a deep learning based surrogate model to determine non-linear aerodynamic characteristics of UAVs. The main advantage of this model is that it can predict the aerodynamic properties of the configurations very quickly by using only geometric configuration parameters without the need for any special input data or pre-process phase. This provides a crucial and explicit design and synthesis tool for mini and small UAVs. To achieve this goal, a large data set, which includes thousands of wing-tail configurations geometry parameters and performance coefficients, was generated using the previously developed and computationally very efficient non-linear lifting line method. This data is used for training the artificial neural network model. The preliminary results show that the neural network model has generalization capability. The aerodynamic model predictions show almost 1-1 coincidence with the numerical data even for configurations with different 2D profiles that are not used in model training. Specifically, the results of test cases are found to capture both the linear and non-linear region of the lift curves, by predicting the maximum lift coefficient, the stall angle of attack, and the characteristics of post-stall region correctly. Similarly, total drag and pitching moment coefficients are predicted successfully. The developed methodology provides the basis for bidirectional design optimization and offers insight for an inverse tool that can calculate geometry parameters for a given design condition.
  • Thumbnail Image
    ItemOpen Access
    A novel physics informed deep learning method for simulation-based modelling
    (AIAA, 2021-01-04) Karali, Hasan; Demirezen, Umut M.; Yukselen, Mahmut A.; Inalhan, Gokhan
    In this paper, we present a brief review of the state of the art physics informed deep learning methodology and examine its applicability, limits, advantages, and disadvantages via several applications. The main advantage of this method is that it can predict the solution of the partial differential equations by using only boundary and initial conditions without the need for any training data or pre-process phase. Using physics informed neural network algorithms, it is possible to solve partial differential equations in many different problems encountered in engineering studies with a low cost and time instead of traditional numerical methodologies. A direct comparison between the initial results of the current model, analytical solutions, and computational fluid dynamics methods shows very good agreement. The proposed methodology provides a crucial basis for solution of more advance partial differential equation systems and offers a new analysis and mathematical modelling tool for aerospace applications