Latent space neural surrogate models have emerged as a powerful class of operator learning approaches, offering strong predictive accuracy, computational efficiency, and architectural flexibility. Despite these advantages, existing approaches often suffer from degradation of the latent physical modes and insufficient capacity to capture dynamic physical interactions, particularly in geometrically complex or physically coupled domains. To address these challenges, we propose the Latent Attention Operator Network (LAON), which integrates a gated physics-adaptive intrinsic mode encoding mechanism with the latent space operator regression. This synergy enables high-fidelity modeling of complex physical field evolution and interactions in arbitrary geometric domains, significantly enhancing the ability to capture localized nonlinear dynamics. We systematically evaluate the LAON in a wide range of PDE benchmarks and challenging scenarios from the real-world automotive industry. Experimental results show that LAON outperforms mainstream attention-based neural PDE surrogate models, achieving over 10% relative improvement in multiple tasks and demonstrating its superior capacity to model the underlying physical characteristics.
Sun et al. (Mon,) studied this question.