Year: 2026
Type: article
Abstract: Predicting the peak wall pressure generated by a collapsing bubble cluster requires expensive resolved simulations and must account for variations in the initial bubble arrangement. A probabilistic surrogate is constructed by combining heteroscedastic Gaussian-process regression with sparse variational inference and a nonlinear autoregressive link between coarse- and fine-grid simulations. Using only descriptors available before collapse, the model estimates the maximum pressure at the wall centre and its predictive uncertainty. The bubble-cluster results demonstrate the importance of local geometry: descriptor enrichment reduces the prediction errors of the extreme wall pressures. Results on the bubble-cluster collapse dataset support the need to model input-dependent residual variance. Comparisons with homoscedastic GPs and alternative probabilistic surrogates demonstrate the benefits of the heteroscedastic formulation through favourable negative log predictive density and continuous ranked probability scores, together with improved predictive interval coverage. Benefits of the multi-fidelity architecture are examined through cost-accuracy analysis with joint sweeping of number of coarse and resolved simulations used for training. Combining 30 high-fidelity and 800 low-fidelity simulations gives slightly lower point-prediction and probabilistic errors than using 50 high-fidelity simulations alone, while reducing training-data generation cost by approximately 32%. These findings support efficient probabilistic wall-load assessment within the tested parameter range and identify local geometric information as a key factor in improving extreme-pressure prediction.
Authors: Jiayang Xu(Conceptualization, Formal analysis, Investigation, Methodology, Validation, Visualization, Writing – original draft, Writing – review & editing), Josef Winter(Conceptualization, Investigation, Methodology, Supervision, Validation, Writing – review & editing), Steffen J. Schmidt(Conceptualization, Funding acquisition, Methodology, Project administration, Resources, Supervision, Writing – review & editing), Nikolaus A. Adams(Conceptualization, Funding acquisition, Methodology, Project administration, Resources, Supervision, Validation, Writing – review & editing)
Institution: Technical University of Munich
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $4,830
Awards: G5144079894, G8603134678
OpenAlex ID: W7203488202
