SCANO: Restoring Boundary-Condition Completeness in Sampled Meshes with a Sliced-Cross-Attention Neural Operator for 3D Thermal Fields

Tianyu Li, Zhiwei Cao, Qingang Zhang, Ruihang Wang, Binyang Song, Yonggang WenarXiv 2026 · arXiv

TL;DR

SCANO is a new neural operator designed to simulate complex 3D physical fields (like heat in a data center) on massive meshes that usually exceed GPU memory. By introducing two plug-and-play modules—Boundary Quadrature Attention and Anchored Global State—it ensures that boundary conditions remain accurate even when the model is trained on small subsets of the data, reducing errors by up to 15-fold.

What Problem Does This Paper Address?

In engineering and physics, we often use "neural operators" to replace slow numerical solvers. These models learn to predict physical fields (like temperature or airflow) almost instantly. However, modern engineering meshes often contain millions of cells. A single training step on a million-cell mesh can consume more memory than a high-end GPU has to offer.

To solve this, researchers usually train models on subsamples (small chunks) of the mesh. The problem is that subsampling is "most damaging at the boundary." If a training sample misses a small but critical boundary patch—like a single server exhaust in a massive data hall—the model fails to learn its impact. Furthermore, standard models often weight boundary points by their count rather than their physical area, leading to biases if the mesh is finer in some places than others. This results in a "sampling gap" where a model trained on samples performs significantly worse than one trained on the full mesh.

Figure 3: Distortions of the boundary measure under sampled training (top row), and how SCANO meets conditions C1–C4 (bottom row, filled tags). The mesh has Dirichlet (black), Neumann (teal) and Robin (red) patches. (a) Unobserved: sampling the boundary nodes (shaded) leaves out the patches that are not sampled. BQA reads all boundary faces in every step (C1). (b) Mis-weighted: weighting by count shifts the unit mass toward finely meshed patches (orange outline). BQA weights the records by face area ay in slice tokens of mass Mg (C2). (c) Attenuated: global modulation (FiLM, adaLN) broadcasts one rank-d code ζ to all points. With BQA, each point attends to the slice tokens with its own weights in every layer (C4). (d) Inconsistent: a global state computed from the sample shifts with the query region R. AGS estimates the state from anchors independent of the sample (pins; C3).

Methods

The authors propose SCANO (Sliced-Cross-Attention Neural Operator), which introduces two backbone-agnostic modules to restore "boundary-condition completeness":

  1. Boundary Quadrature Attention (BQA): Instead of only looking at boundary points that happen to be in the current training sample, BQA reads all boundary faces in every step. It treats the boundary as a continuous integral (quadrature) and weights each face by its actual surface area. These are compressed into "slice tokens" that every point in the mesh can attend to.
  2. Anchored Global State (AGS): Standard models calculate a "global state" based on the current sample, which shifts every time you draw a new sample. AGS instead uses a fixed set of "anchors" drawn independently from the full mesh to keep the global context consistent between training and deployment.
Figure 1: Overview of SCANO and of the consistency conditions C1–C4 that it meets (Section 2.6). (a) Architecture. Boundary Quadrature Attention (BQA) takes all boundary faces as unsampled typed records ϕy = [xy, ny, τy, gy] with areas ay (C1). It pools them into G area-weighted slice tokens zg with masses Mg (bar height, C2). It injects these tokens into every layer of the backbone along a normalised and an unnormalised readout path (C4). The Anchored Global State (AGS) computes the global summary of each layer from volume-weighted anchors. The anchors are drawn independently of the training sample S (C3). (b) Deployment. A model trained on samples is deployed on subsamples (on subgraphs), on a region of interest with its message-passing halo, and on the whole mesh. The whole mesh takes one pass with all cells as anchors. The example is room 1 of Figures 16 and 17.

These modules can be plugged into almost any existing neural operator backbone (like Transolver, GNOT, or MeshGraphNets).

Datasets

The researchers evaluated SCANO on three distinct datasets:

  • Controlled Benchmark: A linear convection-diffusion problem in a unit cube where boundary patches are precisely defined to test specific failure modes.
  • Public Benchmark: A mixed-boundary Poisson problem on a unit disc, used to compare against previous state-of-the-art boundary encoding methods.
  • DCTS (Data Centre Thermal Simulation): A new, industrial-scale suite of 500 different data center rooms, each with roughly a million cells and hundreds of unique boundary patches (servers, cooling units, etc.).

Key Insights and Ideas

  • Attention as Integration: The paper views the attention mechanism as a discretized integral. By using face areas as "quadrature weights," the model becomes invariant to how the mesh is refined. If you split one boundary face into four smaller ones, the model’s output remains identical.
  • The Visibility Bound: The authors mathematically prove that if a boundary patch is entirely missing from a training sample, there is a hard "error floor" that no amount of training can overcome. BQA bypasses this by ensuring no patch is ever "hidden."
  • Consistency is Checkable: The authors derived four conditions (Information Completeness, Mesh Consistency, Sample Consistency, and Response Fidelity) that can be diagnosed before training even begins to predict if a model will fail on large meshes.

Key Results

  • Massive Error Reduction: On the controlled benchmarks, BQA reduced whole-mesh errors by up to 15-fold compared to native backbones.
  • Regional Accuracy: AGS allowed models to predict small regions of interest with the same accuracy as a full-room simulation, cutting regional errors tenfold on public benchmarks.
  • Industrial Performance: On the million-cell DCTS dataset, SCANO matched the accuracy of the best existing models but was 100 times faster than methods using whole-room self-attention.
  • Real-world Utility: SCANO can predict a room-scale thermal field in under one second, making it a viable tool for thermal engineers to test "what-if" scenarios (e.g., "What happens if I double the power to this specific rack?").
Memory usage vs. sample size. SCANO allows training on large meshes that would otherwise exceed GPU memory limits.

Limitations and Future Work

The current study focuses primarily on steady-state problems (where the field doesn't change over time). While the principles of boundary quadrature should apply to time-dependent simulations, this hasn't been tested yet.

Additionally, while the BQA module reads all boundary faces, it still compresses them into a fixed number of tokens ($G$). If a geometry has thousands of tiny, independently varying patches, the "resolution budget" of these tokens might eventually become a bottleneck. The authors suggest that future work could explore hierarchical or multi-scale tokenization to handle even more complex geometries.

Summary updated

Abstract

Training neural operators on million-cell meshes can exceed GPU memory limits. Models are therefore trained on subsamples and deployed on whole meshes or selected regions. However, subsampling can distort boundary information, degrading accuracy. We ask when boundary conditioning under sampling stays consistent with whole-mesh training, and how to aggregate and inject boundary data. Representing the solution as an integral of boundary data against type-dependent measures, we show that subsampling distorts this conditioning in three ways: incomplete boundary coverage or a sample-dependent global state, face weighting by count rather than area, and compression of boundary inputs into a low-dimensional vector. This analysis yields four measurable consistency conditions checkable before training. We satisfy them with two backbone-agnostic modules. Boundary Quadrature Attention (BQA) aggregates all boundary faces, typed and area-weighted, into slice tokens that every query point reads at every layer. Anchored Global State (AGS) estimates the global state from anchors independent of the sampled points. Adding both to a scalable message-passing and slice-attention backbone yields SCANO. We evaluate the approach on controlled and public benchmarks and a data-centre thermal simulation dataset (DCTS) containing million-cell meshes. On both benchmarks, BQA improves every backbone, reducing whole-mesh error up to 15-fold. On all three datasets, AGS aligns regional with whole-mesh predictions, cutting regional error at least tenfold for most backbones on the public benchmark. Together, the modules nearly halve the best original backbone's whole-mesh error on both benchmarks. SCANO predicts a room-scale field in under one second, over 100 times faster than whole-room anchor self-attention.
BibTeX
@misc{li2026scano,
  author = {Tianyu Li and Zhiwei Cao and Qingang Zhang and Ruihang Wang and Binyang Song and Yonggang Wen},
  title = {{SCANO: Restoring Boundary-Condition Completeness in Sampled Meshes with a Sliced-Cross-Attention Neural Operator for 3D Thermal Fields}},
  year = {2026},
  eprint = {2611.04603},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.04603},
}
Figure