ADGB-ONet: an attention-driven gradient balancing operator network for scattering pressure field prediction

A computational study in Applied Intelligence reports that an attention-based neural network for scattered pressure field prediction achieved a mean relative error (MRE) of 2.88% in comparisons with other deep-learning models. The proposed Attention-Driven Gradient Balancing Operator Network, or ADGB-ONet, produced the lowest prediction error among the compared approaches. Its MRE was 23.4% lower than that of OFormer, identified in the study as the strongest baseline by this measure, and 58.0% lower than the MRE of the baseline MIONet. The results concern computational prediction of pressure fields used in acoustic metasurface design; the available report does not describe physical experiments or establish performance outside the tested benchmark.

The study addresses a computational design problem: finite element simulations can be expensive, making exploration of large acoustic metasurface design spaces impractical, while existing deep-learning models may struggle to reconstruct localized structures with high spatial gradients. ADGB-ONet is designed to improve that reconstruction through a progressive representation strategy combining spectral enrichment, condition-dependent feature interactions and spatially adaptive reconstruction. The network does not require explicit gradient information. Instead, it learns how to adjust its reconstruction sensitivity from spatial features and contextual information, allowing different parts of a predicted field to be handled according to the variation present in the input representation.

The first component is a Fourier Embedding layer, which maps raw spatial coordinates into sine and cosine bases at multiple scales. The authors describe this as enriching the coordinate representation and improving the trunk network’s ability to resolve rapid spatial variations in scattered pressure fields. A Multi-Branch Attention Mechanism then couples these enriched spatial features with frequency and geometric descriptors. In place of fixed fusion rules, the mechanism uses condition-guided feature interactions, meaning that the combination of features is guided by information associated with the input location and its context. This architecture is intended to make the network more responsive to localized field structures rather than treating every region with the same fixed processing rule.

For reconstruction, the Adaptive Attention Reconstruction Module, or AARM, processes the modulated features through multiple branches using different transformation scales. A Gate Neural Network generates coordinate-dependent weights that aggregate the branch outputs adaptively. In regions where variation differs across the field, the gate can therefore give different branches different influence, producing spatially adaptive reconstruction. The model also uses linear attention to capture spatial dependencies among neighboring sampling points. Ablation studies reported in the study found that AARM made the largest individual contribution to performance among the tested architectural components. Taken together, the Fourier features, guided attention, multiscale branches and learned spatial gating form the mechanism proposed to improve reconstruction of localized high-gradient structures. The study presents ADGB-ONet as a reliable surrogate model for accelerating acoustic metasurface design, but the supplied abstract does not quantify computational savings, provide the domain size or sample counts, or report physical validation.

Subject of Research: Neural surrogate modeling of scattered acoustic pressure fields for acoustic metasurface design

Article Title: ADGB-ONet: an attention-driven gradient balancing operator network for scattering pressure field prediction

Article References: Tan, A., Wu, J., Sheng, Y., Zhang, R., & Zhou, Q. (2026). ADGB-ONet: an attention-driven gradient balancing operator network for scattering pressure field prediction. Applied Intelligence, 56(16), Article 510. https://doi.org/10.1007/s10489-026-07538-x

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07538-x

Keywords: ADGB-ONet, scattered pressure fields, acoustic metasurfaces, deep learning, neural surrogate models, operator learning, attention mechanisms, Fourier features, computational acoustics, attention-driven, gradient, balancing