Machine learning methods for building reduced-order models

Building constructions, buildings and structures
Authors:
Matveeva, A.I. Kravchinsky, S.A. Avdonyushkin, D.V. Novokshenov, Aleksei Dmitrievich Borovkov, Aleksey Ivanovich Karchevskaia, Anna Stanislavovna Sherbakov, Sergei Sergeevich Klimkovich, Nikita Mikhailovich Podgayskaya, Daria Aleksandrovna
Abstract:

The article explores methods for building surrogate models using neural networks and Gaussian process regression. It also discusses creating reduced-order models through singular value decomposition. These methods are tested on a technical engineering object, where reduced models are developed from hydrodynamic simulations. The article highlights the use of the Levenberg-Marquardt algorithm for training neural networks. This algorithm performs better than gradient descent and shows high accuracy with real data. Additionally, Gaussian process regression with the Matеrn kernel offers precise interpolation, making it effective for solving complex engineering problems. Singular value decomposition combined with Gaussian process regression allows to obtain a state vector that accurately represents the model in seconds.

***ARTICLE IN PRESS***