Statistically-informed surrogate models combining linear dimension reduction and neural networks for blast wave propagation
Year: 2026
Type: article
Abstract: In this paper, we introduce a framework to solve complex multi-block vector-to-vector regression problems. The proposed approach combines supervised or unsupervised linear dimension reduction with nonlinear mapping in latent space using neural networks. A subspace error metric is employed to determine the optimal reduced space dimensionality given the training set size. To learn the mapping in the reduced space, we propose two novel statistically-informed strategies: a variance-informed method that trains a single neural network, and a dependency-informed method that constructs a sequence of neural networks tailored to each output coordinate. Our methods are applied to a blast wave propagation problem which consists in predicting the maximum overpressure with two sources of uncertainty, namely the height of burst and the elevation field both scaled by the explosion yield. Numerical experiments across various training set sizes and reduced space dimensionalities demonstrate the accuracy and robustness of the proposed methods. Furthermore, the resulting surrogate model enables efficient statistical analysis of the predictions, including conditional probability distributions and global sensitivity fields.
Source: Journal of Computational Science
Language: English
Subfield: Statistical and Nonlinear Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,560
OpenAlex ID: W7149334697