Technical Summary: COMPOL – A Scalable Neural Operator Framework for Multi-Physics Simulations

1. Problem Statement

Multi-physics simulations, governed by coupled partial differential equations (PDEs), are essential for modeling complex interactions in fields ranging from fluid mechanics to environmental engineering. Traditional numerical methods (e.g., finite element, finite difference) face significant computational bottlenecks when addressing high-dimensional or tightly coupled multi-physics problems. While recent advances in neural operators (such as the Fourier Neural Operator, FNO) have improved computational efficiency by approximating solution mappings between infinite-dimensional function spaces, existing methods often fail to capture the intricate, non-linear correlations inherent in coupled physical processes.

Current neural operator frameworks typically treat multiple physical processes independently or employ simplistic coupling strategies (e.g., feature concatenation). These approaches inadequately represent the dynamic interdependencies that span multiple spatial and temporal scales, leading to insufficient predictive accuracy for complex systems like fluid-structure interactions, chemical reaction-diffusion systems, and multiphase geological flows.

2. Methodology: COMPOL Framework

To address these limitations, the authors propose COMPOL (Coupled Multi-Physics Operator Learning), a novel framework designed to explicitly model interactions among multiple physical processes within a shared latent space.

Core Architecture

COMPOL is architecture-agnostic but is primarily instantiated using Fourier Neural Operator (FNO) layers. The framework processes $M$ interacting physical systems in parallel through the following stages:

1. Latent Embedding: Initial physical input functions $f^m(x)$ for each process $m$ are mapped into a shared latent feature space via process-specific linear projections $P^m$.
2. Iterative Latent Transformation: The framework employs $L$ layers of coupled neural operators. In each layer $l$, the latent representation $v^m_{l-1}$ for process $m$ is transformed by an independent neural operator $h^m_l$. Crucially, this transformation is augmented by a global interaction feature $z_{l-1}$.
3. Attention-Based Aggregation: The global interaction feature $z_{l-1}$ is computed using a sophisticated attention mechanism. The latent states from all $M$ processes are projected into Query ($Q$), Key ($K$), and Value ($A$) representations. The aggregation state is derived as a weighted sum of value vectors, where weights $\alpha_j$ dynamically quantify the significance of interactions involving the $j$-th process. This allows the model to explicitly capture non-linear cross-process dependencies at every layer.
4. Reconstruction: After $L$ iterations, dedicated linear projections $Q^m$ reconstruct the latent representations back into the final physical solution spaces $g^m(x)$.

Complexity and Scalability

The computational complexity is dominated by the Fourier operations ($O(LM d_h \cdot N \log N)$), with the attention aggregation contributing $O(LM^2 d_h)$. The framework is designed to be scalable, supporting an arbitrary number of physical processes while maintaining efficiency in typical scenarios where spatial discretization points $N$ are large.

3. Key Contributions

The paper outlines three primary contributions:

1. Versatile Framework: A coupled multi-physics operator learning framework that is architecture-agnostic and enhances traditional operator layers with sophisticated feature aggregation tailored for coupled systems.
2. Attention-Based Latent Aggregation: A novel technique that robustly captures dynamic inter-process dependencies across an arbitrary number of physical processes, separating process-specific evolution from global interaction modeling.
3. Empirical Validation: Extensive demonstration of COMPOL's effectiveness and scalability across diverse benchmarks, showing consistent improvements over state-of-the-art methods.

4. Experimental Results

The authors evaluated COMPOL on five diverse coupled multi-physics PDE benchmarks:

• 1-D Lotka-Volterra: Predator-prey dynamics.
• 1-D Belousov-Zhabotinsky: Chemical oscillations with three reactants.
• 2-D Gray-Scott: Pattern formation in chemical reactions.
• 2-D Multiphase Flow: Oil-water flow in porous media.
• 2-D Thermo-Hydro-Mechanical (THM): Coupling of pore pressure, temperature, and strain.

Performance Metrics:

• Predictive Accuracy: Across all benchmarks and training set sizes (512 and 1024 samples), COMPOL consistently achieved the lowest relative $L_2$ prediction errors compared to baselines (FNO, UFNO, Transolver, DeepONet, and CMWNO). Improvements ranged up to 72.41% over baselines with 512 samples and 63.64% with 1024 samples.
• Ablation Studies:
  • Capacity vs. Coupling: When FNO variants were scaled to match COMPOL's parameter count (by increasing modes, width, or depth), they still underperformed COMPOL. This confirms that the performance gain stems from the explicit cross-process coupling mechanism rather than increased model capacity.
  • Single-Process Capability: Even when applied to a single-process problem (Burgers' equation), COMPOL outperformed standard baselines, suggesting its aggregation mechanisms provide beneficial regularization and feature selection even without explicit multi-physics coupling.
  • Spectral Analysis: COMPOL demonstrated superior agreement with ground-truth spectral energy distributions, indicating better fidelity in capturing scale-wise dynamics.

5. Significance and Claims

The paper claims that COMPOL establishes a new standard for learning the complex dynamics of coupled multi-physics systems. By explicitly modeling interdependencies within latent spaces, it overcomes the "insufficient representation of inter-process dynamics" found in current neural operator frameworks. The authors position COMPOL as a scalable, effective solution that bridges the gap between the efficiency of neural operators and the physical fidelity required for complex multi-physics simulations.

Limitations Acknowledged:
The authors modestly note that current experiments rely on fixed structured grids, leaving generalization across spatial resolutions unexplored. Additionally, the evaluation focuses on single-step predictions (initial to final state), and the framework's stability under long-term autoregressive rollouts (where error accumulation is a risk) remains a direction for future work.

Reproducibility:
The authors have made the code and data publicly available to support open science and reproducibility.