Abstract:Reduced-order models (ROMs) become increasingly popular in industrial design and optimization processes, since they allow to approximate expensive high fidelity computational fluid dynamics (CFD) simulations in near real-time. The quality of ROM predictions highly depends on the placement samples in the spanned parameter space. Adaptive sampling strategies allow to identify regions of interest, which feature e.g. nonlinear responses with respect to the parameters, and therefore enable the sensible placement of… Show more
“…In view of this, the low prediction accuracy when using only images as input data is not only a consequence of the non-singularity of the solution but also of the nature of this data-driven method, the number and location of the input data in the parameter space, as reported in the literature. [23] When considering the Kriging interpolation model with both images and numerical data for solving inverse tasks, the system behavior was fully captured as shown in the parity plots for 6 outputs in Figure 7.…”
Section: Resultsmentioning
confidence: 99%
“…In view of this, the low prediction accuracy when using only images as input data is not only a consequence of the non‐singularity of the solution but also of the nature of this data‐driven method, the number and location of the input data in the parameter space, as reported in the literature. [ 23 ]…”
Development of the Vertical Growth Freeze crystal growth process is a typical example of solving the ill‐posed inverse problem, which violates one or more of Hadamard's well‐posedness criteria of solution existence, uniqueness, and stability. In this study, different data‐driven approaches are used to solve inverse problems: Reduced Order Modelling method of Proper Orthogonal Decomposition with Inverse Distance weighting (ROM POD InvD), an approximation method of Kriging and Artificial Neural Networks (ANN) employing images, combination of images and numerical data and solely numerical data, respectively. The ≈200 training data are generated by Computational Fluid Dynamics (CFD) simulations of the forward problem. Numerical input data are related to the temperatures and coordinates in 10 characteristic monitoring points in the melt and crystal, while the image input data are related to the interface shape and position. Using the random mean squared error as a criterion, the Kriging method based on images and numerical data and the ANN method based on numerical data are able to capture the system behavior more accurately, in contrast to the ROM POD InvD method, which is based solely on images.
“…In view of this, the low prediction accuracy when using only images as input data is not only a consequence of the non-singularity of the solution but also of the nature of this data-driven method, the number and location of the input data in the parameter space, as reported in the literature. [23] When considering the Kriging interpolation model with both images and numerical data for solving inverse tasks, the system behavior was fully captured as shown in the parity plots for 6 outputs in Figure 7.…”
Section: Resultsmentioning
confidence: 99%
“…In view of this, the low prediction accuracy when using only images as input data is not only a consequence of the non‐singularity of the solution but also of the nature of this data‐driven method, the number and location of the input data in the parameter space, as reported in the literature. [ 23 ]…”
Development of the Vertical Growth Freeze crystal growth process is a typical example of solving the ill‐posed inverse problem, which violates one or more of Hadamard's well‐posedness criteria of solution existence, uniqueness, and stability. In this study, different data‐driven approaches are used to solve inverse problems: Reduced Order Modelling method of Proper Orthogonal Decomposition with Inverse Distance weighting (ROM POD InvD), an approximation method of Kriging and Artificial Neural Networks (ANN) employing images, combination of images and numerical data and solely numerical data, respectively. The ≈200 training data are generated by Computational Fluid Dynamics (CFD) simulations of the forward problem. Numerical input data are related to the temperatures and coordinates in 10 characteristic monitoring points in the melt and crystal, while the image input data are related to the interface shape and position. Using the random mean squared error as a criterion, the Kriging method based on images and numerical data and the ANN method based on numerical data are able to capture the system behavior more accurately, in contrast to the ROM POD InvD method, which is based solely on images.
“…Our algorithm is also applicable to both non‐intrusive and projection‐based ROMs, whereas previous works have focused exclusively on non‐intrusive ROMs. A wide range of adaptive sampling algorithms is compared in a work 11 by Karcher and Franz. When applied to two external aerodynamics problems, our results show that the proposed adaptive sampling algorithm outperforms using LHS alone in predicting physical fields and integral quantities of lift and drag.…”
The use of reduced‐order models (ROMs) in physics‐based modeling and simulation is a popular tool for drastically lowering the computational cost associated with high‐fidelity simulations. ROMs use training data from a set of computed high‐fidelity simulations with different design parameters that control physical and geometric properties of the full‐order model. The quality of the training data dictates the performance of the ROM, making the choice of training design parameters important. A widely used method for generating training parameters for ROMs is Latin hypercube sampling (LHS), a statistical method that aims to maximize the distance and minimize the correlation amongst produced samples. However, LHS fails to account for the physics of the full‐order model and can lead to an over‐representation of some physical regimes in the training data while neglecting others. In this work, we present a computationally efficient adaptive sampling method for ROMs using Isomap, a versatile algorithm for nonlinear dimensionality reduction. Using an initial number of samples, the adaptive sampling algorithm iteratively generates samples based on a low‐dimensional manifold of the training data. When applied to two external aerodynamics problems, the proposed adaptive sampling algorithm offers significantly improved performance in predicting physical fields and coefficients of lift and drag for a given computational budget for both non‐intrusive and projection‐based ROMs.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.