Wednesday, April 12, 2023 10:00AM

Master Thesis Proposal

Felix Luo

(Advisor: Prof. Suresh Menon)


"Evaluation of Convolutional Neural Networks for Modeling Blast Propagation in
Multi-room Bunkers"


Wednesday, April 12 at
10:00 a.m.
Montgomery Knight Building 32
5

 

Abstract

Machine learning and neural networks have revolutionized high tech fields by enhancing the ability of predictive models to determine complex nonlinear relationships in seemingly uncorrelated data. This increase in capability of predictive models has not excluded the aerospace field with neural network models being developed for closure models for Reynolds Averaged Navier Stokes simulations, sub grid turbulence contributions to Large Eddy Simulations, flow reconstruction from low dimensional or low-resolution data, and controls. However, past work in integrating neural networks with aerospace applications have typically focused on steady state flows with well-defined turbulence statistics whereas very little work has been done in the field of non-steady flows with local geometrical dependencies such as in the case of blasts in multi-room geometries. Therefore, this study seeks to extend the capabilities of neural networks into this gap in capabilities to provide predictive abilities for key features such as the peak pressures in blasts in multi-room bunkers. Traditionally, models predicting the response of blasts in multi-room bunkers require intricate understanding of blast dynamics to develop a computationally efficient model in simplistic geometries or the use of costly computational fluid dynamics code in more geometrically complex bunkers. However, a neural network strikes a balance between model design simplicity and computational efficiency.  Furthermore, since blast dynamics in multi-room bunkers are largely dependent on local geometrical dependencies such as walls and corners, a convolutional neural network based model provides a generalizable model with lower training cost over traditional feed forward neural networks. Consequently, this thesis proposes to apply a convolutional neural network (CNN) to predict the peak pressures in blasts in multi-room bunkers. To achieve this goal, the proposed thesis will comprise of quantifying advantages of a CNN over traditional feed forward network for this application, determination of geometrical complexity extrapolative ability for a CNN based model, and characterization of the CNN sensitivity to varying bunker sizes. Through the completion of these objectives, a computationally efficient CNN based model for the prediction of peak pressures for blasts in multi-room bunkers can be developed to allow for rapid iterative design of safer bunker designs. 

Committee


• Prof. Suresh Menon – School of Aerospace Engineering (advisor)
• Prof. Lakshmi Sankar – School of Aerospace Engineering
• Prof. Spencer Bryngelson – School of Computational Science and Engineering