GLOBE: A New Architecture for Physics-Informed Machine Learning in Engineering

In the high-stakes world of engineering, from aerospace design to automotive innovation, computational fluid dynamics (CFD) simulations are indispensable but notoriously slow, often consuming days or weeks of supercomputer time. This bottleneck has spurred interest in machine learning (ML) surrogates that can approximate these simulations faster, yet many existing models fail to respect fundamental physical laws, leading to poor generalization in real-world scenarios. A new paper introduces GLOBE, a neural surrogate architecture that draws inspiration from boundary-element s and equivariant ML to address these shortcomings, promising substantial gains in accuracy and practicality for industrial applications like computer-aided engineering (CAE). By structurally enforcing translation, rotation, and parity equivariance, along with discretization and units invariance, GLOBE aims to bridge the gap between data-driven flexibility and physical rigor, potentially revolutionizing how engineers approach complex simulations.

GLOBE's architecture is built on a synthesis of domain-inspired insights, representing PDE solutions as superpositions of learnable, Green's-function-like kernels evaluated from boundary faces to target points. The model operates across three hierarchical levels: at the bottom, kernel functions compute influences in an all-to-all manner between sources and targets, ensuring a global receptive field critical for elliptic PDEs like those in aerodynamics. These kernels are designed with inductive biases from physics, such as far-field decay envelopes that stabilize extrapolation and Padé-approximant MLPs that mimic the rational-function behavior of Green's functions. At the middle level, multiscale composition allows the model to handle multiple characteristic length scales by summing outputs from parallel kernel branches, each tuned to different reference lengths derived from the governing equations. This multiscale approach enables efficient representation across disparate spatial scales, from viscous boundary layers to geometric features, enhancing the model's ability to capture complex flow phenomena.

The top level of GLOBE incorporates communication hyperlayers that propagate latent information between boundary partitions before final evaluation, mediating long-range coupling akin to traditional boundary-element s but with lower computational cost. This hyperlayer stack computes source strengths and exchanges data across different boundary condition types, such as no-slip or inlet boundaries, ensuring that the model respects problem-specific constraints without relying on dense linear solves. The architecture's design ensures key mathematical properties: translation-equivariance through the use of relative positions, rotation-equivariance via invariant feature engineering and local vector reprojection, and discretization-invariance through area-weighted aggregation that converges to a continuous boundary integral in the fine-mesh limit. Additionally, rigorous nondimensionalization throughout the data pipeline guarantees units-invariance, making the model robust to changes in physical units and aligning with engineering standards.

In evaluations on the AirFRANS dataset, which includes steady incompressible RANS solutions over NACA airfoils, GLOBE demonstrated remarkable performance improvements. On the Full split, it reduced mean-squared error by roughly 200 times compared to the dataset's reference baselines and about 50 times relative to the next-best model, enf2enf. In low-data scenarios like the Scarce split, it achieved over 100 times lower error on velocity and pressure fields and over 600 times lower error on surface pressure than Transolver, another state-of-the-art model. Qualitative showed sharp near-wall gradients, coherent wake structures, and limited errors under extrapolation in Reynolds number and angle of attack, highlighting its generalizability. The model's compact size of 117,000 parameters and ability to evaluate fields at arbitrary points during inference further underscore its practicality for industrial deployment, where computational efficiency and accuracy are paramount.

Beyond numerical accuracy, GLOBE's robustness to non-watertight meshes has strong for real-world CAE workflows, where geometric imperfections often cause traditional solvers to fail. In stress tests, the model maintained functionality even when trained on severely decimated boundary meshes with 50% face removal and topological defects, degrading gracefully rather than catastrophically. This suggests that GLOBE can handle CAD-derived geometries with minor imperfections, reducing preprocessing overhead and accelerating design cycles. The paper argues that such physics-informed architectures are essential for building trust in ML surrogates among engineers, as they adhere to exact physical constraints like conservation laws and symmetries, unlike black-box models that may learn spurious correlations. Future work could extend GLOBE to 3D aerodynamic datasets and other PDE domains, leveraging its scalable design for broader industrial impact.

Reference: Sharpe, P. (2025). GLOBE: Accurate and Generalizable PDE Surrogates Using Domain-Inspired Architectures and Equivariances. arXiv:2511.15856v1.