Fast Multi-Objective Aerodynamic Optimization Using Sequential Domain Patching and Multifidelity Models

Probabilistic Neural Data Fusion for Learning from an Arbitrary Number of Multi-Fidelity Data Sets

“…The performance of an NN is highly sensitive to its architecture and hyperparameters if the training data is small, unbalanced, and multi-fidelity. To reduce this sensitivity and leverage the low costs of training a single NN on small data, we perform automated hyperparameter tuning 11 . To this end, we use RayTune [55] and Hyperopt to find the optimum hyperparameters and architecture by minimizing the five-fold crossvalidation errors on predicting the high-fidelity data.…”

“…Early works in this field focused primarily on hierarchically linking bi-fidelity data. For instance, in space mapping [8][9][10] or multi-level [11][12][13] techniques the inputs of the LF data are mapped following formulations such as x l = F (x h ) where x l and x h are the inputs of LF and HF sources, respectively. In this equation, F (•) is a transformation function whose predefined functional form is calibrated such that y l (F (x h )) approximates y h (x h ) as closely as possible.…”

“…In this equation, F (•) is a transformation function whose predefined functional form is calibrated such that y l (F (x h )) approximates y h (x h ) as closely as possible. These techniques are useful in applications where higher fidelity data are obtained by successively refining the discretization in simulations [11,12], e.g., by refining the mesh when modeling the flow around an airfoil or estimating the fracture toughness of a microstructure. The main disadvantages of space mapping techniques are that (1) they rely on iterative and time-consuming analysis for choosing a near-optimal functional form for F (•), (2) they cannot jointly fuse more than two data sources at a time, (3) they quantify similarity/discrepancy between the sources based on pre-defined functions whose space may not include the true discrepancy, and (4) they do not quantify some uncertainty sources (such as lack of data) and are rarely formulated within a Bayesian setting that leverages prior information.…”

On the multi-fidelity approach in surrogate-based multidisciplinary design optimisation of high-aspect-ratio wing aircraft

“…Examples of surrogate based aeroelastic instability parametric search [44] and aerodynamic damping estimation for flutter calculations [45,46] can be found in the literature. Multi-fidelity methods were also used to estimate flutter boundary, gust response [47] and aerodynamic optimisation [48,49], although there is very scarce literature in multi-fidelity models applied to MDO considering flutter [50].…”

Rapid Multi-Criterial Antenna Optimization by Means of Pareto Front Triangulation and Interpolative Design Predictors

“…The techniques mentioned in the previous paragraphs are stochastic in the sense that at some stage of the optimization process, randomized search procedures (in particular, natureinspired methods) are employed, either to handle the problem objectives directly or at the level of the surrogate model. Recently, several deterministic surrogate-assisted MO techniques have been proposed including point-by-point Pareto front exploration [53], a bisection method [54], as well as sequential domain patching (SDP) [55]. Although these algorithms have been developed to handle twoobjective problems, some generalized versions have been proposed as well (generalized bisection [56], or generalized SDP [57]).…”