Learning Dynamical Coupled Operator For High-dimensional Black-box Partial Differential Equations

Yichi Wang, Tian Huang, Dandan Huang, Zhaohai Bai, Xuan Wang, Lin Ma, Haodi ZhangIJCAI 2025 · DOI

TL;DR

This paper introduces the Dynamical Coupled Operator (DCO), a novel neural operator designed to tackle high-dimensional, black-box partial differential equations (PDEs). DCO improves upon existing methods by incorporating temporal dynamics and a more robust input/query encoding scheme, leading to significantly better performance and stability in complex PDE problems.

What Problem Does This Paper Address?

Solving complex partial differential equations (PDEs) is crucial in many scientific and engineering fields, but traditional numerical methods are often computationally expensive and struggle with scalability. Deep operator networks (DONs) have emerged as a promising alternative, learning mappings between function spaces to approximate PDE solutions. However, existing DONs face challenges:

  • Information Bottlenecks: Mapping low-dimensional inputs to high-dimensional outputs can lead to information loss and unstable training.
  • Limited Encoding: Current methods often fail to adequately represent input functions and query locations, hindering generalization.
  • Lack of Guarantees: Some DON variants sacrifice theoretical guarantees of universal approximation, impacting their reliability.

This paper aims to address these limitations by proposing a more stable, robust, and theoretically sound operator learning framework for high-dimensional, black-box PDEs.

Methods

The core of this work is the Dynamical Coupled Operator (DCO), which introduces several key innovations:

  • Dynamically Coupled Functions: DCO learns a coupled function that incorporates temporal dynamics. This approach aims to reduce information loss and improve training robustness by better capturing the interplay between input functions and query positions.
  • Adaptive Spectral Input Function Encoder: To handle non-stationary input functions, DCO employs Empirical Mode Decomposition (EMD). EMD decomposes signals into intrinsic mode functions (IMFs) that capture oscillatory components at different scales, providing a more adaptive and effective feature representation than traditional spectral methods.
  • Hybrid Query Location Encoding: DCO combines a normalized Radial Basis Function (RBF) kernel with discrete Fourier encoding. This hybrid approach captures both localized spatial information and periodic characteristics of query locations, enhancing the model's ability to generalize to unseen query points.
  • Theoretical Guarantees: The authors provide theoretical proofs demonstrating that DCO satisfies universal approximation properties, ensuring its applicability to a wide range of PDE problems.

Datasets

The paper evaluates DCO on the ERA5 dataset from WeatherBench. This dataset contains governing black-box PDEs with a dimensionality of $d_u = 54$ for input functions and $d_s = 5$ for output functions. The input functions include ground temperature (t2m), atmospheric temperature (t), geopotential (z), and ground wind vector (u10, v10). The experiments are conducted at three different scales and for various lead times (3, 6, 9, and 12 hours).

Key Insights and Ideas

  • Temporal Dynamics are Crucial: Incorporating temporal dynamics into the operator learning process is key to overcoming information bottlenecks and improving stability, especially for dynamic systems.
  • EMD for Adaptive Feature Extraction: Empirical Mode Decomposition proves to be a more flexible and effective method for extracting features from non-stationary input functions compared to fixed spectral methods.
  • Hybrid Encoding for Robustness: Combining localized (RBF) and periodic (Fourier) encoding for query locations leads to better generalization and performance on unseen query points.
  • Theoretical Foundation Matters: Providing universal approximation guarantees instills confidence in the model's applicability to a broad range of complex problems.

Key Results

DCO significantly outperforms existing state-of-the-art DONs and other operator methods on the ERA5 dataset.

  • Improved Accuracy: DCO achieves lower Mean Squared Error (MSE) and Root Mean Squared Error (RMSE) across all tested lead times and physical quantities compared to DeepONet, LOCA, and IPOT. For example, at a 3-hour lead time, DCO's average MSE is 0.0619, outperforming DeepONet (0.0750) and IPOT (0.0643).
  • Enhanced Stability: The paper highlights that the coupled (u, y) representation in DCO is more stable than decoupled approaches, as shown in Figure 1.
  • Superior Generalization: DCO demonstrates excellent performance in predicting unseen query locations, outperforming baselines in both MSE and RMSE.
  • Ablation Study: Removing components like the adaptive spectral encoder (w/o A) or the hybrid query location encoding (w/o B) leads to a noticeable drop in performance, underscoring the importance of these novel contributions.

Limitations and Future Work

  • Limitations: The paper does not explicitly detail limitations beyond the inherent challenges of high-dimensional PDE solving. The ablation study suggests that while DCO is robust, further improvements might be possible.
  • Future Work: The authors express a desire to develop a unified and universally guaranteed operator architecture based on their work. This suggests a direction towards more generalizable and theoretically robust operator learning frameworks.

Summary updated

BibTeX
@inproceedings{wang2025learning,
  author = {Yichi Wang and Tian Huang and Dandan Huang and Zhaohai Bai and Xuan Wang and Lin Ma and Haodi Zhang},
  title = {{Learning Dynamical Coupled Operator For High-dimensional Black-box Partial Differential Equations}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/1044},
}
Figure