Uncertainty quantification using Auto-tuned Surrogates of CFD model Simulating Supersonic flow over tactical missile body

In contemporary practices, Computational Fluid Dynamics (CFD) based tools are increasingly applied to build high fidelity First Principles based Models (FPMs) for designing tactical missile systems. However, optimization, sensitivity analysis and uncertainty quantification using such models still remain to be extremely tedious and, hence, are performed offline. Artificial Neural Networks (ANNs) are known to be efficient machine learning based models capable of modelling large amount of nonlinearities in the data. However, heuristics involved in their modelling prevent their application as surrogate models to computationally intensive FPMs. In this work, a novel algorithm, aimed at simultaneous optimal estimation of architecture (number of hidden layers and nodes in each layer), training sample size and activation function in ANNs is proposed. The proposed algorithm classifies as a generic multi-objective evolutionary neural architecture search strategy to design ANNs. It is solved using the population based evolutionary optimization algorithm called Nondominated Sorting Genetic Algorithm-II. The high fidelity data to train the ANNs are obtained from CFD model for simulating supersonic flow over a tactical missile body in ANSYS Fluent^©. The obtained ANNs with a test set accuracy of around 99% are then used for quantifying uncertainties, due to order of discretization, flux and Spalart-Allmaras turbulence model in CFD simulations, on the coefficients of lift, drag and rolling moment using the method of analysis of variance (ANOVA). The applicability of optimally designed MLP surrogates for performing ANOVA based study presents novel insights into the epistemic uncertainties arising due to the CFD model which can help in realistic designs of tactical missiles.