Point Queries in Neural Surrogates

• Point queries are evaluations of neural fields at arbitrary spatial or spatiotemporal coordinates, enabling mesh-independent and high-resolution inference.
• They employ cross-attention decoders and latent bottlenecks to achieve linear scaling, which supports efficient evaluation in industrial-scale simulations.
• Point queries decouple field representation from fixed discretization, facilitating real-time multi-fidelity workflows in automotive, aerospace, and multi-physics applications.

Point queries are a fundamental operational mode for neural surrogate models and operator learning frameworks in computational physics, especially in Universal Physics Transformer (UPT) architectures. The “point query” paradigm refers to the evaluation of a learned field, surrogate, or operator at arbitrary user-specified spatial (and sometimes temporal) coordinates, rather than being restricted to a fixed grid or data layout determined a priori by the training mesh or simulation discretization. This capability enables efficient memory scaling, flexibility across input/output geometries, and high-resolution output generation, and it is central to the practical deployment of modern neural operator approaches for industrial-scale simulations and scientific computing.

1. Definition and Mathematical Role of Point Queries

A point query is the process of evaluating a neural field or operator at a set of spatial or spatiotemporal coordinates $\{y_i\},\ i=1,\ldots,M$. Given a latent representation $z$ encoding the physical state or simulation input, and a decoder $\mathcal{D}$, the point query operation computes

$\hat u(y_i) = \mathcal{D}(z, y_i)$

for any set $\{y_i\}$, where $y_i$ need not be associated with the training mesh. The decoder is typically conditionally parameterized on the latent code $z$ output by an encoder acting on initial/boundary conditions, mesh data, or geometric representations ($z = \mathcal{E}(x)$). The pointwise evaluation can be parallelized for arbitrary $M$ and supports flexible inference, including variable-resolution and mesh-free output (Alkin et al., 2024, Alkin et al., 17 Oct 2025).

Point queries are essential for “coordinate-based” neural surrogates, enabling outputs to be generated in a continuous domain without entangling the computational cost to the density of the training mesh, as opposed to grid-tied or array-based models.

2. Architectures Supporting Point Queries

Several UPT and neural operator architectures are explicitly constructed to enable point query capability:

• Cross-attention decoders: UPTs use cross-attention mechanisms wherein each query coordinate $y_i$ is embedded and attends to a latent set $z$0 (or a set of anchor tokens in AB-UPT), producing outputs $z$1. No self-attention is computed among the queries, so inference cost scales linearly with $z$2 (Alkin et al., 2024, Alkin et al., 17 Oct 2025).
• Latent bottlenecks: By compressing the input field/mesh to a small set of latent tokens, the decoder need only attend from queries to these latents, decoupling model size and compute from the number of output coordinates.
• Supernode and anchor-based decoders: Geometry-preserving approaches apply supernode pooling or anchor point selection to reduce attention complexity, supporting efficient batched evaluation at arbitrary queries (Alkin et al., 13 Feb 2025, Alkin et al., 17 Oct 2025).
• Conditional neural fields: Many architectures (e.g., field decoders in GP-UPT) embed each query $z$3 via sinusoidal or MLP positional embeddings before cross-attending to the latent, permitting continuous output evaluation (Alkin et al., 13 Feb 2025).

Table 1 summarizes typical mechanisms enabling point queries in neural operator frameworks.

Architecture	Mechanism	Point Query Complexity
UPT core (Alkin et al., 2024)	Cross-attention to latent tokens	$z$4
AB-UPT (Alkin et al., 17 Oct 2025)	Cross-attention to surface/volume anchors	$z$5
GP-UPT (Alkin et al., 13 Feb 2025)	Cross-attention to supernodes	$z$6
PINN, FNO	Tied to global grid	$z$7 (must query whole grid)

3. Computational and Scalability Implications

Point querying enables substantial scalability advantages over conventional full-grid surrogates:

• Linear scaling of inference and memory: Decoupling the number of inference points from model size or latent representation allows large output fields (volumetric, surface) to be evaluated flexibly. UPT-based models have achieved inference on up to 45 million points in <35 seconds on a single GPU by leveraging anchor-based decoding (Alkin et al., 17 Oct 2025).
• Mesh-independence: The capacity to predict fields at arbitrary coordinates eliminates the requirement to match training and inference meshes. This allows, e.g., evaluation on user-provided CAD meshes, solution-adapted meshes, or adaptive point clouds not seen during training (Alkin et al., 17 Oct 2025).
• Arbitrary-resolution outputs: By evaluating outputs at any desired set of coordinates—including dense grids for high-resolution visualization or sparse samples for steering optimization—the model supports multi-fidelity workflows without retraining (Alkin et al., 2024).

This paradigm is distinct from standard convolutional or grid-structured models (FNO, U-Net), whose output is locked to the discretization on which the model was trained and which exhibit quadratic or cubic scaling with resolution (Alkin et al., 2024).

4. Theoretical and Practical Significance

Point query-capable neural surrogates possess several methodological and theoretical advantages:

• Unified surrogate operators: Point queries generalize to unstructured, hybrid, or dynamic meshes, supporting Eulerian and Lagrangian physical systems with the same architecture (Alkin et al., 2024).
• Efficient surrogate design for industry: In industrial CFD, surface and volume queries at tens of millions of points are supported with constant model memory via anchor- or latent-based cross-attention (Alkin et al., 17 Oct 2025).
• Decoupling field representation from discretization: Outputs need not share the discretization of the training data, enabling direct comparison to arbitrary ground truths, meshless visualization, or integration with other modeling tools (e.g., force integration on user-supplied surfaces).
• Horizon-independent rollout: In latent-operator UPTs, the ability to query the field at arbitrary time and/or space after evolving the latent forward enhances the flexibility of temporal and spatial extrapolation (Alkin et al., 2024).

A plausible implication is that effective generalization to novel geometries, mesh refinements, and new boundary instantiations—core to many engineering and scientific workflows—is directly enabled by point querying.

5. Comparative Results and Ablation Studies

Ablative comparisons document the impact of point querying:

• GP-UPT's “field decoder” enables high-resolution output querying but slightly increases MSE versus a “point decoder” (restricted to training supernode locations). Loss in resolution-locked accuracy is offset by flexibility and scalability in arbitrary-point evaluation (Alkin et al., 13 Feb 2025).
• AB-UPT achieves near-perfect prediction of integrated surface and volume quantities (drag/lift) even when evaluated on out-of-distribution CAD test meshes by using point query decoders anchored to reference points (Alkin et al., 17 Oct 2025).

Table 2 exemplifies point-query-based inference runtimes for industry-scale CFD surrogates using AB-UPT (Alkin et al., 17 Oct 2025):

Task (SHIFT-Wing)	#Points	Inference Time (H100)
CAD surface	200K	0.6 s
CFD surface	3M	8.4 s
Full volume	6M	16.8 s

6. Applications and Limitations

Point querying is centrally used in the following contexts:

• Automotive/aerospace CFD surrogates: Real-time field, surface, and force prediction on full-vehicle meshes (Alkin et al., 17 Oct 2025).
• Multi-physics and foundation models: Transferrable neural solvers supporting arbitrary domain geometries (Alkin et al., 2024, Wiesner et al., 17 Sep 2025).
• Adaptive and hybrid meshing: Downstream tasks such as design optimization, high-resolution post-processing, and uncertainty quantification.

However, some limitations and subtleties are also present:

• Models may incur a minor loss in per-point accuracy versus architectures restricted to fixed predictor locations—but gain substantially in generalization and flexibility (Alkin et al., 13 Feb 2025).
• Query cost, while linear in $z$8, may become non-negligible for very dense output meshes unless anchoring or batching strategies are used (Alkin et al., 17 Oct 2025).
• Input features and positional embeddings must be designed to handle out-of-distribution locations for true mesh-independent generalization.

7. Outlook and Research Directions

The ubiquity and proven value of point queries in neural operator-based surrogate modeling has established them as a fundamental abstraction for universal physics simulation, enabling industry-scale deployment. Ongoing research aims to further reduce cross-attention complexity, optimize anchor selection, extend to irregular and hybrid domains, and improve adaptive sampling strategies for uncertainty quantification and active learning (Alkin et al., 17 Oct 2025, Alkin et al., 2024, Zhou et al., 2024).

A plausible implication is that the flexibility and scalability delivered by point querying will become standard in future foundation models for physics, supporting real-time simulation, surrogate optimization, and continuous-domain field generation across scientific and engineering domains.