The Daniel Guggenheim School of Aerospace Engineering at Georgia Tech



The ASM initiative is exploring advanced surrogate modeling. The primary goal of the ASM initiative is to advance the capability of surrogate modeling so that surrogates can be used with more accuracy on a wider range of modeling types, including dynamic models, stochastic models, and highly non-linear models. Surrogate types being researched include neural networks (NN), Gaussian Processes, and Kriging Models. Current research topics for this area include dynamic surrogates, surrogates of stochastic models, and NN learning algorithms. In addition, selection of the correct design of experiments for a given surrogate and development of new types of design of experiments are being researched.

Research Questions

  • When creating surrogates of stochastic models, how can a surrogate of the distribution of the results be created, and how do you choose which distribution is most appropriate?
  • What improvements can be made to Neural Network Learning Algorithms?
  • How can we better create surrogates of dynamic models?
  • Can we find a way to more intelligently (or adaptively) choose the samples taken during a design of experiments to reduce the number of cases required to create accurate surrogate models?

Summary of Developed Techniques


Developed by: Woong Je Sung

The modeling and simulation (M&S) tasks in aeronautics are increasingly demanding highfidelity, physics-based analysis tools to fulfill ever-increasing technological challenges. These challenges are being derived from the need for the intelligent and adaptive aeronautical systems greatly improving the performance and robustness of aircraft and the air transportation system as a whole. But the integration of high-fidelity, physics-based analysis modules such as computational fluid dynamics (CFD) tools into the multidisciplinary design optimization (MDO) methodologies is only practical within the restricted design space because of the prohibitive requirement for the computational resources as the design freedom increases. The artificial neural network (ANN) is, recently, extending its application area to the fluid mechanics field showing promising prediction capability for the highly complex, nonlinear, unsteady flow phenomena using relatively small amount of training data. The surrogate modeling using the ANN has shown the fundamental advantage compared to other efficiency enhancing approaches in the MDO paradigm such as adjoint method and multi-fidelity modeling. But the lack of the proper mean to the architectural selection of the ANN greatly hinders the systematic assessment for the generalization efficiency and, hence, it’s further enhancement. The main hypothesis of this study is that proper architectural selection for the ANN can significantly enhance the generalization capability of the ANN and, hence, this might enable the global optimization without the degradation of analysis fidelity or the sacrifice of the design freedom. Contrast to the conventional connection-centric encoding schemes, this study proposes a novel neuron-centric genetic encoding scheme for representation of the architecture and the initial weights of the ANN, based on the Extended Neuron Model, which extends the conventional simple neuron model to the information processing unit capable of establishment of its own connections. With additional proposition of the Neuronal Constraint as a useful architectural constraint for the ANN development, a novel neuroevolution methodology, GENESIS (GEnetic NeuroEvolution by Synapse-Induced Structure) has been developed. For the given training data, GENESIS evolves both the structure and the weight parameters of the ANN using genetic algorithm (GA) as an evolutionary framework and the back-propagation-Levenberg-Marquardt algorithm as a network training method. As a proof-of-concept case, one particular fluid mechanics problem has been modeled and the results are compared with conventional multi-layer perceptron (MLP)'s results.