.png){kind=icon}
Return
A Hierarchical Interaction-Aware Framework for Sample-Efficient Surrogate Modeling
.png){kind=icon}
PRE
AI
DOI：10.1109/tii.2026.3737080.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
This article presents a hierarchical, interaction-aware data-generation framework that improves the sample efficiency of surrogate model training by prioritizing samples according to the physical importance of design variables and their interactions. Surrogate-model performance depends strongly on the quality of training data, while engineering systems often provide prior knowledge about the relative importance of design variables. The proposed framework exploits this knowledge to organize data generation hierarchically, enabling neural networks to identify dominant variables and interaction structures at an early stage of learning. By exposing physically meaningful structures directly through the training data, the framework reduces the dependence of initial feature discovery on network architecture. Experiments on an induction-coil design problem demonstrate that the proposed approach substantially improves sample efficiency. While conventional random sampling requires approximately 70% of the available data to achieve over 90% predictive accuracy, the proposed framework achieves comparable accuracy using only 30% of the data. Moreover, the performance gap between a 4-layer MLP and a 17-layer ResNet is substantially reduced, indicating that the proposed sampling strategy alleviates architectural dependence during feature discovery. These results demonstrate that incorporating physical interaction hierarchy into data generation can improve surrogate-model learning under limited data.
Keywords:
Data sampling
digital twin
interaction hierarchy
neural networks (NNs)
sample efficiency
surrogate modeling
Journal
IF:
9.9
Papers:
8.6K
Citations:
6.0W
Organization
Cited Papers
No cited papers available