betaFeedback

WorksComputer Science

Adaptive Coordinate Transforms for Neural Operators

Abstract

Neural operators have achieved promising performance on partial differential equations (PDEs), but most existing models are built on fixed Eulerian coordinates. This mismatch between evolving physical structures and static coordinates creates spatial misalignment, leading to unnecessarily non-local operator mappings and reinforcing a smoothness preference near sharp transitions. Inspired by adaptive coordinate transformations in classical PDE analysis, we propose the Adaptive Coordinate Transform (ACT) block, a plug-and-play module for data-driven geometric adaptation in neural operators. ACT blocks resolve this structural limitation by learning adaptive coordinate systems within the operator learning pipeline. Specifically, given an input feature, the ACT block learns a coordinate transformation and represents the same feature under the transformed coordinates via differentiable sampling. This operation preserves the underlying signal while changing its spatial representation, equivalent to expressing the same physical quantity in different coordinate systems. By adapting the coordinate system to the data, ACT allows the network to better track evolving structures, reduce operator complexity, and dynamically focus on critical features to improve learning. We evaluate the proposed approach across diverse PDE benchmarks and multiple neural operator architectures. Experimental results demonstrate consistent and significant improvements in predictive accuracy, indicating that learning coordinate systems provides a powerful mechanism for enhancing operator learning.

Introduction

Neural operators have emerged as a powerful framework for solving partial differential equations (PDEs) [15]. By learning mappings between input and output fields, they can serve as efficient surrogates for classical numerical solvers. Architectures such as Fourier Neural Operators (FNO) [16], ConvNeXt-UNet (CNextU) [18], and Transolver [29] have shown strong performance in modeling complex physical dynamics.

Despite their success, a fundamental limitation exists: most existing neural operators are built upon fixed Eulerian coordinates. This static coordinate system is inherently misaligned with the evolving physical structures (e.g., propagating waves, moving vortices) that characterize many PDEs. This spatial misalignment leads to highly non-local and complex input-output mappings, significantly increasing learning complexity and hindering generalization [11]. Furthermore, these operators exhibit a secondary challenge: a smoothness preference [24]. Constrained

Preprint. by their inherent inductive biases, they tend to favor overly smooth mappings and struggle to maintain the correct regularity of the solution, often blurring high-frequency physical phenomena like shocks or sharp transitions.

From the perspective of classical PDE analysis, these challenges of complexity and regularity are primarily consequences of using fixed Eulerian coordinates. Rather than solving such evolving dynamics within a static coordinate system, classical methods often employ adaptive coordinate transformations, ranging from Arbitrary Lagrangian-Eulerian (ALE) formulations [10] to moving mesh methods [12], to dynamically reshape the computational coordinates. By allowing the coordinate system to track advecting structures and locally compress around sharp gradients, these techniques simultaneously reduce spatial misalignment and avoid excessive numerical smoothing. Consequently, they unravel complex spatiotemporal evolution into a smoother, more localized, and better resolved form.

Inspired by these classical principles of adaptive coordinate transformation, we propose the Adaptive Coordinate Transform (ACT) block, a plug-and-play module designed to equip neural operators with data-driven geometric realignment capabilities. By predicting a coordinate adjustment field from the current feature map, the ACT block resamples intermediate features via differentiable sampling and integrates them through a residual connection. Specifically, we interleave ACT blocks layer-wise within foundational operator backbones (e.g., FNO, CNextU, and Transolver), creating an architecture capable of continuous geometric realignment.

Crucially, we uncover an autonomous role decoupling mechanism within this layer-wise architecture. Based on their topological locations, the ACT blocks dynamically divide their labor to mimic classical adaptive techniques. Intermediate blocks act as progressive spatial trackers, realigning evolving features to neutralize spatial misalignment and maintain a stable latent space. Conversely, the final ACT block acts as a regularity modulator. By implicitly compressing the local coordinate space, it leverages the Jacobian multiplier to explicitly amplify spatial gradients, effectively counteracting the smoothness preference and recovering sharp physical transitions.

In summary, our work bridges the gap between classical adaptive mesh techniques and modern operator learning. The main contributions are as follows:

  • We propose the Adaptive Coordinate Transform (ACT) block, a learnable geometric realignment module inspired by classic PDE techniques, specifically designed to overcome spatial misalignment and the smoothness preference of neural operators built on fixed Eulerian coordinates.

  • Extensive experiments across diverse foundational architectures (FNO, CNextU, and Transolver) demonstrate that ACT-augmented models significantly improve predictive accuracy and better capture complex dynamics.

  • We reveal an autonomous role decoupling phenomenon in layer-wise networks: intermediate ACTs track advection for latent alignment, while the terminal ACT reconstructs sharp gradients via Jacobian-driven spatial compression.

Related Work

Neural Operators on Fixed Coordinates Neural operators learn mappings between function spaces and have achieved strong performance on PDEs. Representative paradigms include spectral operators such as the Fourier Neural Operator (FNO) [16], encoder-decoder baselines built on modern convolutional blocks [18], and transformer-based architectures such as Transolver [29]. These models offer distinct inductive biases for spectral mixing, locality-aware hierarchical feature extraction, and long-range interaction modeling, respectively. However, these approaches generally assume a fixed Eulerian coordinate system and do not explicitly model coordinate transformations. For dynamics involving transport and advection, they address spatial misalignment only indirectly through spectral global coupling, stacked local convolutions, or attention-based global interactions, so the operator remains unnecessarily non-local in the original coordinates.

Coordinate Adaptation in PDE Solvers and PDE Learning Reducing complexity through adaptive coordinate transformations or mesh adaptation is standard in PDE computation. In learningbased settings, some works learn deformation maps or adaptive meshes to improve discretization and solver efficiency [25, 30, 11], while others use learned mappings to parameterize irregular domains or canonicalize data before operator learning [17, 26]. These approaches are related in spirit, but they mainly target mesh optimization, domain parameterization, or preprocessing rather than layer-wise geometric realignment of latent operator representations.

Learnable Deformation Modules Learnable deformation is also well established in deep learning, especially in image registration [5, 2], deformable convolution [4], and deformable attention [31]. These methods motivate ACT architecturally, especially through local offset prediction and differentiable sampling. However, they are mostly designed for vision tasks, and even deformation-based neural operator variants outside PDE learning [21] remain application-specific. In contrast, we treat adaptive coordinate transformation as a native, layer-wise component of neural operator learning, using geometric realignment throughout the backbone to reduce spatial misalignment and simplify the learned operator.

Method

This section presents our framework for learning neural operators under adaptive coordinate transformations. We first describe the spatial misalignment induced by fixed Eulerian coordinates, then introduce the proposed ACT block, and finally show how it is integrated with a neural operator backbone.

Problem Setup

Let Ω⊂Rd\Omega\subset\mathbb{R}^{d} be a bounded spatial domain equipped with a fixed Eulerian coordinate system. Let U(Ω×[0,T];Rdu)U(\Omega\times[0,T];\mathbb{R}^{d_u}) and V(Ω×[0,T];Rdv)V(\Omega\times[0,T];\mathbb{R}^{d_v}) denote the input and output function spaces, respectively. We consider an operator F\mathcal{F} that maps functions from UU to VV as:

F:U→V,u(x,t)↦v(x,t)=F[u](x,t),(1)\mathcal{F}: U \to V,\quad u(x,t) \mapsto v(x,t) = \mathcal{F}[u](x,t), \tag*{(1)}

which is typically induced by an underlying PDE. Here u∈Uu \in U and v∈Vv \in V are spatiotemporal functions evaluated at a given point x∈Ωx \in\Omega and time t∈[0,T]t \in[0,T]. Depending on the task, u(x,t)u(x,t) may represent the initial and boundary conditions or a forcing term, while v(x,t)v(x,t) typically represents the solution field at a future time step.

Neural operators approximate F\mathcal{F} by a parameterized network Fθ\mathcal{F}_{\theta}, which is typically formulated as:

Fθ=Q∘GθL−1∘⋯∘Gθ0∘P,(2)\mathcal{F}_{\theta} = Q \circ\mathcal{G}_{\theta}^{L-1} \circ\cdots\circ\mathcal{G}_{\theta}^{0} \circ P, \tag*{(2)}

where PP pointwise lifts the input function to a latent feature space, Gθl\mathcal{G}_{\theta}^{l} denotes the ll-th latent operator update (e.g., a spatial integral operator using Fourier or attention mechanisms) acting on the entire domain, and QQ is a local projection mapping the final hidden representation back to the physical output space at xx.

Spatial Misalignment and Geometric Realignment

The main difficulty arises from the mismatch between evolving physical structures and fixed Eulerian coordinates. Physical dynamics are inherently characterized by continuous transport and diffusion, causing structures (e.g., a propagating wave or a moving vortex) to naturally advect and spread across the spatial domain over time.

When such dynamics are modeled on a fixed Eulerian coordinate system, the fundamental mismatch between the static coordinates and evolving structures breaks this spatial alignment, resulting in a highly non-local and complex input-output mapping. For the neural operator Fθ:u(x,t)↦v(x,t)\mathcal{F}_{\theta}: u(x,t) \mapsto v(x,t) in Eq. (2), this means that a large part of the apparent complexity may stem from spatial misalignment in the representation rather than from the underlying dynamics themselves.

However, classical PDE analysis provides a clear solution to this problem: rather than solving such evolving dynamics within a static coordinate system, classical methods often employ coordinate transformations to simplify the problem. Concretely, we can define a coordinate transformation

ψ:Ω→Ω,x↦ξ:=ψ(x),(3)\psi: \Omega\to\Omega,\quad x \mapsto\xi:= \psi(x), \tag*{(3)}

that maps the original Eulerian coordinates to a new coordinate system that inherently tracks the evolving structures. Then the neural operator can be reformulated under the new coordinates as:

Fˉθ:uˉ(ξ,t)↦vˉ(ξ,t),(4)\bar{\mathcal{F}}_\theta: \bar{u}(\xi,t) \mapsto\bar{v}(\xi,t), \tag*{(4)}

where uˉ(ξ,t)=u(ϕ(ξ),t)\bar{u}(\xi,t) = u\left(\phi(\xi),t\right), vˉ(ξ,t)=v(ϕ(ξ),t)\bar{v}(\xi,t) = v\left(\phi(\xi),t\right) and ϕ:=ψ−1\phi:= \psi^{-1} denotes the inverse map. Therefore, the transformed operator Fˉθ\bar{\mathcal{F}}_\theta may admit a simpler representation, since adaptive coordinate transformations can reduce spatial misalignment and reorganize complex spatiotemporal evolution into a smoother and more localized form [10, 12].

Adaptive Coordinate Transform (ACT) Block

These observations suggest that geometric realignment should be handled explicitly rather than absorbed implicitly by the operator backbone. Following this idea, we introduce an adaptive coordinate transform (ACT) block, which uses the current feature map to realign sampling locations before the next operator update. Motivated by this perspective, ACT realizes geometric realignment in a learnable, layer-wise manner through feature-conditioned local sampling maps, rather than by explicitly learning a globally invertible coordinate transform.

Here, we implement ACT in a residual and compositional form, inspired by the deformable attention mechanism [31]. Rather than predicting a completely new coordinate system, each layer estimates an incremental coordinate offset relative to the identity mapping. Formally, given a latent feature map hh defined over Ω\Omega, the block predicts a collection of head-specific coordinate offsets {Δh(m)}m=1M\{\Delta_h^{(m)}\}_{m=1}^{M}. These offsets define sampling maps through the residual formulation

ϕh(m)(ξ)=ξ+Δh(m)(ξ),ξ∈Ω,m=1,…,M,(5)\phi_h^{(m)}(\xi) = \xi+ \Delta_h^{(m)}(\xi),\quad\xi\in\Omega,\quad m = 1,\ldots,M, \tag*{(5)}

where Δh(m)\Delta_h^{(m)} is conditioned on the current feature map. Under this residual parameterization, Δh(m)≡0\Delta_h^{(m)} \equiv0 recovers the identity sampling pattern. Therefore, by controlling the magnitude of the offsets, we can make the realignment easier to control and stabilize.

The overall architecture is illustrated in Figure 1, which can be decomposed into three main steps: coordinate offset prediction, grid resampling, and feature merging.

Figure 1. Architecture of the ACT-enhanced operator model.

Step 1: Coordinate Offset Prediction. We first project the input feature map hh and split it into MM independent value tensors {Vm}m=1M\{V_m\}_{m=1}^{M} for head-specific offset prediction.

To incorporate positional information, the normalized base coordinates ξ\xi are concatenated with each value tensor along the channel dimension, forming joint representations fin(m)=[Vm,ξ]f_{\mathrm{in}}^{(m)} = [V_m,\xi]. Instead of using separate predictors for each head, we employ a single offset predictor that processes all heads in parallel. The predictor consists of a lightweight multi-scale convolutional module:

Δh,raw(m)=Conv⁡fuse(σ([Conv⁡1×1(fin(m)),Conv⁡3×3(fin(m)),Conv⁡dilated(fin(m))])),(6)\Delta_{h,\mathrm{raw}}^{(m)} = \operatorname{Conv}_{\mathrm{fuse}}\left(\sigma\left([\operatorname{Conv}_{1 \times1}(f_{\mathrm{in}}^{(m)}), \operatorname{Conv}_{3 \times3}(f_{\mathrm{in}}^{(m)}), \operatorname{Conv}_{\mathrm{dilated}}(f_{\mathrm{in}}^{(m)})]\right)\right), \tag*{(6)}

where σ\sigma denotes the GELU activation. The predictor outputs a dd-dimensional offset field that represents coordinate displacements. To stabilize the realignment, we bounding the raw offsets by a maximum displacement magnitude α\alpha as

Δh(m)=αtanh⁡(Δh,raw(m)).(7)\Delta_h^{(m)} = \alpha\tanh\left(\Delta_{h,\mathrm{raw}}^{(m)}\right). \tag*{(7)}

Step 2: Grid Resampling. Using the predicted offsets, we define sampling maps via Eq. (5). These maps specify, for each output location ξ\xi, the corresponding sampling location ϕh(m)(ξ)\phi_h^{(m)}(\xi) in the original coordinates. The feature representation under the transformed coordinates is obtained by

V~m(ξ)=Vm(ϕh(m)(ξ)).(8)\widetilde{V}_m(\xi) = V_m\left(\phi_h^{(m)}(\xi)\right). \tag*{(8)}

This operation preserves the underlying feature values while altering the sampling locations, resulting in a different spatial representation of the same feature.

To handle out-of-domain coordinates, we impose physically meaningful boundary conditions. Standard differentiable sampling operators typically handle such coordinates by clamping them to the nearest border. To better respect the underlying physical topology, our ACT block instead maps out-of-bounds sampling locations back into the valid domain prior to the actual grid sampling. For example, in a 1D periodic domain Ω=[−1,1]\Omega= [-1, 1], an out-of-bounds sampling coordinate ϕ(ξ)\phi(\xi) is mapped back into the domain via:

ϕ^(ξ)=((ϕ(ξ)+1) mod 2)−1.(9)\hat{\phi}(\xi) = ((\phi(\xi) + 1) \bmod2) - 1. \tag*{(9)}

This coordinate extension ensures that subsequent sampling queries the physically consistent locations across periodic boundaries, rather than simply replicating edge values. Our experimental results demonstrate that this physics-informed coordinate conditioning prevents boundary artifacts and yields significant performance improvements.

Step 3: Feature Merging. The sampled features V~m\widetilde{V}_m are concatenated and projected via a pointwise linear layer WoW_o. In parallel, a residual branch preserves the original representation via hres=Wreshh_{\mathrm{res}} = W_{\mathrm{res}}h. The final output is

h~=Tθ(h)=Norm⁡(hres+Wo(Concat⁡(V~1,…,V~M))).(10)\widetilde{h} = \mathcal{T}_{\theta}(h) = \operatorname{Norm}\left(h_{\mathrm{res}} + W_o\left(\operatorname{Concat}(\widetilde{V}_1,\ldots,\widetilde{V}_M)\right)\right). \tag*{(10)}

This residual formulation allows the ACT block to introduce coordinate-level adjustments while retaining the original feature representation, ensuring stable and incremental updates.

Instantiation with Neural Operator Backbone

To prevent spatial misalignment from re-emerging as features propagate through depth, geometric realignment is introduced in a layer-wise manner rather than only once at the input or output. Denoting the operator update at layer ll by Gθl\mathcal{G}_{\theta}^{l} and the corresponding adaptive coordinate block by Tθl\mathcal{T}_{\theta}^{l}, the hidden features evolve as

hl+1=Tθl∘Gθl(hl),l=0,…,L−1.(11)h^{l+1} = \mathcal{T}_{\theta}^{l} \circ\mathcal{G}_{\theta}^{l}(h^l), \qquad l = 0,\ldots,L-1. \tag*{(11)}

Accordingly, the overall network takes the form

Fθ=Q∘TθL−1∘GθL−1∘⋯∘Tθ0∘Gθ0∘P.(12)\mathcal{F}_{\theta} = \mathcal{Q} \circ\mathcal{T}_{\theta}^{L-1} \circ\mathcal{G}_{\theta}^{L-1} \circ\cdots\circ\mathcal{T}_{\theta}^{0} \circ\mathcal{G}_{\theta}^{0} \circ\mathcal{P}. \tag*{(12)}

In our implementation, each latent operator block is followed by an ACT block, so the hidden stack alternates between operator propagation and geometric realignment. The lifting and projection operators P\mathcal{P} and Q\mathcal{Q} are standard pointwise layers; further details on resampling under learned coordinates and regularization are deferred to Appendix § 8.

Experiments

This section first describes the experimental setting, including the datasets, baselines, evaluation metric, and ACT-specific design choices, and then presents the main evaluation and ablation results.

Experimental Setting

Dataset Description: To comprehensively evaluate model performance across both two-dimensional and three-dimensional settings, we conduct experiments on six representative datasets drawn from three widely used sources. Specifically, Navier_Stokes follows the benchmark setting introduced in FNO [16]; Diff_Reaction and Shallow_Water are taken from PDEBench [27]; and Active_Matter, TRL_2D, and MHD_64 (a 3D dataset) are selected from The Well [23]. This diverse suite allows us to assess model robustness across different spatial dimensions and physical regimes under consistent benchmark protocols. In all cases, we aim to learn an operator Fθ:u(⋅,t0),…,u(⋅,tk)↦u(⋅,tk+1)F_{\theta}: u(\cdot,t_0),\ldots,u(\cdot,t_k) \mapsto u(\cdot,t_{k+1}), which maps a history window of k+1k+1 observed states to the following state. Detailed dataset statistics are summarized in Table 3, and full dataset descriptions are provided in Appendix § 9.

DatasetCSResolutionn_stepsTrainValidTest
Navier_Stokes(x,y)(x,y)128×128128 \times128504005050
Diff_Reaction(x,y)(x,y)128×128128 \times128101800100100
Shallow_Water(x,y)(x,y)128×128128 \times128101800100100
Active_Matter(x,y)(x,y)256×256256 \times256811752426
TRL_2D(x,y)(x,y)128×384128 \times3841017299
MHD_64(x,y,z)(x,y,z)64364^3100771010

Table 3. Detailed description of the datasets, including coordinate system (CS), spatial resolution, number of time-steps per trajectory (n_steps), and the number of trajectories in the training, validation, and test sets.

Baseline Models: To rigorously evaluate the potential of our proposed ACT mechanism, we integrate it into three representative state-of-the-art baselines spanning distinct paradigms: CNextU [18] (a cutting-edge convolutional architecture widely adopted in image processing), Transolver++ [29, 19] (an attention-based operator), and FNO [16] (a classic spectral operator). By applying ACT to these diverse baseline models under identical experimental protocols, we ensure a fair and systematic assessment of its universal effectiveness and modeling enhancements.

Metrics: We evaluate model performance using the Normalized Root Mean Squared Error (NRMSE), which measures the discrepancy between the predicted field xx and the ground-truth field yy while being invariant to the scale of the target field. The metric is defined as

NRMSE⁡(x,y)=∑i=1N∣xi−yi∣2∑i=1N∣yi∣2+ε,(13)\operatorname{NRMSE}(x,y)=\sqrt{\frac{\sum_{i=1}^{N}|x_i-y_i|^2}{\sum_{i=1}^{N}|y_i|^2+\varepsilon}}, \tag*{(13)}

where xix_i and yiy_i denote the predicted and ground-truth values at the ii-th spatial location, respectively, NN is the total number of spatial grid points, and ε=10−7\varepsilon=10^{-7} is a small constant introduced for numerical stability.

Our Design: Building upon the baseline models, we augment each backbone with ACT blocks that perform data-driven coordinate adaptation. Each ACT block introduces only a modest number of additional parameters and incurs very limited overhead for large operator backbones such as FNO and CNextU. These modules are interleaved with the backbone layers so that intermediate features are repeatedly represented in adaptive coordinate systems, while preserving the original inductive biases of each baseline architecture. For hierarchical architectures such as CNextU, where the feature-map resolution changes across depth, we use a depth-adaptive maximum displacement magnitude (i.e., α\alpha in Eq. (7)), that scales with the local grid spacing. Detailed architectural and training configurations are deferred to Appendix § 10. Finally, to ensure a fair and rigorous evaluation, all other model settings—including network depth, hidden width, training protocols, and optimization hyperparameters—remain strictly identical to those of the corresponding vanilla baselines.

Evaluation Results

Benchmark Comparison: As shown in Table 1, we compare the inference performance of three representative backbones (CNextU, Transolver, and FNO), with and without the proposed ACT module, across six PDE datasets. ACT consistently reduces NRMSE across all backbone-dataset pairs in this benchmark. These improvements hold across distinct architectural families, including the convolutional CNextU, the attention-based Transolver, and the spectral FNO. The average relative error reductions reach 76.4% for Transolver, 38.4% for FNO, and 30.8% for CNextU. Overall, these results show that ACT is a broadly effective, model-agnostic module for improving neural operator accuracy across diverse physical systems. For more comprehensive visualization results, please refer to Appendix § 12.

DatasetCNextUTransolverFNO
VAN+ACTVAN+ACTVAN+ACT
Navier_Stokes6.13%4.96%↓19.1%32.29%1.35%↓95.8%2.22%0.62%↓72.1%
Diff_Reaction4.03 ‰1.20 ‰↓70.2%2.20 ‰0.69 ‰↓68.6%1.66 ‰1.34 ‰↓19.3%
Shallow_Water2.11 ‱1.58 ‱↓25.1%3.49 ‱0.78 ‱↓77.7%1.38 ‱0.93 ‱↓32.6%
Active_Matter3.43%2.31%↓32.7%18.61%4.91%↓73.6%4.49%2.69%↓40.1%
TRL_2D9.14%7.57%↓17.2%20.34%6.98%↓65.7%11.97%8.95%↓25.2%
MHD_6412.08%9.63%↓20.3%29.30%6.82%↓76.7%15.40%9.09%↓41.0%
Average ↓—↓30.8%—↓76.4%—↓38.4%

Table 1. Model Performance Comparison. NRMSE are reported. Gray columns correspond to models augmented with ACT. Values marked with ↓ indicate the relative error reduction compared to the vanilla (VAN) setting. The symbols ‰ and ‱ denote parts per thousand (10−310^{-3}) and parts per ten thousand (10−410^{-4}), respectively. The best results are highlight in Bold.

Ablation Study: We conduct a controlled ablation study on the Navier–Stokes equation using FNO as the backbone to disentangle three factors: increased model capacity, a single terminal realignment, and layer-wise realignment throughout the network. The results are summarized in Table 2. Specifically, we compare four variants: (i) the vanilla FNO (around 16.8M parameters), (ii) a larger FNO with approximately twice the number of parameters (around 37.9M), (iii) FNO augmented with only one ACT block at the last layer, and (iv) FNO with ACT blocks applied at every layer. This comparison isolates whether the observed gains arise merely from more parameters or from where ACT is inserted.

MethodFNOFNO (2× params)FNO + Final ACTFNO + Layer-wise ACT
NRMSE2.22%1.80%1.15%0.62%

Table 2. Ablation study on the Navier–Stokes equation using FNO as the backbone. Introducing adaptive coordinate transformation (ACT) modules significantly improves performance compared to increasing model capacity.

Figure 2. Predicted density ρ\rho results of TRL_2D dataset for different models.

The results show that simply increasing model capacity yields only limited improvement. In contrast, introducing only a final ACT block already produces a substantially larger gain. Applying ACT blocks in a layer-wise manner yields a further improvement over both the larger-capacity baseline and the final-ACT-only variant. This controlled comparison indicates that the performance gains primarily stem from the introduction of coordinate adaptation, rather than merely increasing model capacity.

Insights

ACT delivers strong gains across architectures and datasets, which naturally raises a mechanistic question: what exactly are the ACT blocks doing inside the network? To better understand this, we visualize the layer-wise feature evolution and the effect of adding only the terminal ACT block in

Figure 3. These visualizations suggest that ACT does not play a uniform role across depth. Instead, different blocks appear to divide their labor: earlier blocks mainly reorganize the latent representation, whereas the terminal block is more directly tied to the recovery of sharp local structure. Below, we unpack these two roles in turn.

Figure 3. Visualization of feature representations induced by ACTs. (i) Layer-wise visualization of intermediate features, showing that the effect of ACT varies across depth, with both the first and the terminal blocks producing clear realignment. (ii) Features and gradient norms before and after adding only the final ACT block, showing that even this isolated terminal realignment sharpens local structure. Together, the two panels provide qualitative evidence that ACT first simplifies the latent operator mapping and then uses the final block to recover sharp detail.

How Intermediate ACT Blocks Simplify the Latent Operator

Earlier ACT blocks primarily serve as geometric realignment modules: instead of forcing every latent operator update to model transport or deformation on a fixed grid, they absorb part of this burden into learned coordinate adaptation, so that subsequent operator updates act on hidden representations with more regular spatial structure. In this way, they mitigate spatial misalignment before it accumulates across depth.

To illustrate this effect, we consider a simple setting of an integral operator:

F:u↦v,v(x)=∫ΩK(x,y) u(y) dy.(14)\mathcal{F}: u \mapsto v,\qquad v(x) = \int_{\Omega} K(x,y)\,u(y)\,\mathrm{d}y. \tag*{(14)}

Next, we examine how input-side and output-side coordinate transformations modify the kernel representation of this operator. Detailed derivations are deferred to Appendix § 11.

Input-side transformation. For an input-side coordinate transformation written as y=ϕin(ξ)y = \phi_{\mathrm{in}}(\xi), define the transformed input by uˉ(ξ)=u(ϕin(ξ))\bar{u}(\xi) = u(\phi_{\mathrm{in}}(\xi)). The corresponding effective kernel is

K~in(x,ξ)=K(x,ϕin(ξ))∣det⁡∇ϕin(ξ)∣,(15)\widetilde{K}_{\mathrm{in}}(x,\xi) = K(x,\phi_{\mathrm{in}}(\xi))\left|\det\nabla\phi_{\mathrm{in}}(\xi)\right|, \tag*{(15)}

Thus, input-side transformations re-parameterize the integration variable and change how information is aggregated from the input domain.

Output-side transformation. For an output-side coordinate transformation written as x=ϕout(ξ)x = \phi_{\mathrm{out}}(\xi), define the transformed output by vˉ(ξ)=v(ϕout(ξ))\bar{v}(\xi) = v(\phi_{\mathrm{out}}(\xi)). The effective kernel becomes

K~out(ξ,y)=K(ϕout(ξ),y).(16)\widetilde{K}_{\mathrm{out}}(\xi,y) = K(\phi_{\mathrm{out}}(\xi),y). \tag*{(16)}

Hence, the transformation re-parameterizes the evaluation variable and changes where the operator is queried.

Summary. Input-side and output-side transformations therefore modify the operator in complementary ways: one changes the aggregation coordinates, while the other changes the query coordinates. In both cases, the original operator FF is rewritten in transformed coordinates as an equivalent operator with a modified kernel representation. This is the mechanism by which geometric realignment can simplify the operator seen by subsequent latent updates.

How the Final ACT Block Restores Sharp Local Structure

Once earlier ACT blocks have reduced spatial misalignment in the latent features, the final block can focus on restoring sharp local structure near the output. This is consistent with the visualization above that ACT exhibits a depth-dependent division of labor: earlier ACT blocks mainly simplify the latent operator mapping, whereas the last block is important for reconstructing local detail.

Panel (ii) of Figure 3 provides a more direct check of this idea by isolating the case in which only the final ACT block is added. Even in this restricted setting, the final block visibly sharpens the feature representation and its gradient structure, providing qualitative evidence that terminal realignment contributes directly to recovering sharp local transitions.

One plausible mechanism for this sharpening effect is local coordinate compression. If the final ACT block induces an output-side re-parameterization vˉ(ξ)=v(ϕ(ξ))\bar{v}(\xi) = v(\phi(\xi)), then the chain rule gives

∇ξvˉ(ξ)=∇ϕ(ξ)⊤∇xv(ϕ(ξ)).(17)\nabla_{\xi}\bar{v}(\xi) = \nabla_{\phi}(\xi)^{\top}\nabla_{x}v(\phi(\xi)). \tag*{(17)}

Therefore, when ϕ\phi locally compresses coordinates along a certain direction, the Jacobian matrix ∇ϕ(ξ)\nabla\phi(\xi) acts as a multiplier on the physical gradient and can increase the magnitude of the corresponding derivative in the transformed representation. This provides a simple explanation for why the terminal ACT block can recover sharper local transitions.

The ablation results in Table 2 further support this interpretation. Adding only a final ACT already improves performance, while layer-wise ACT yields further gains beyond both the vanilla backbone and a larger-capacity baseline. Together, these results suggest that ACT distributes its effect across depth: intermediate blocks reduce geometric complexity, whereas the final block is especially important for recovering local sharpness.

Limitations

Despite these encouraging results, several aspects of ACT deserve further discussion. First, both the benchmark results and the analyses suggest that ACT is most beneficial when spatial misalignment, transport, mixing, deformation, or sharp moving interfaces are major sources of difficulty. In such settings, geometric realignment can simplify the latent operator while the terminal ACT block helps recover local sharpness. By contrast, for problems whose dynamics are already well represented in fixed Eulerian coordinates or do not exhibit pronounced geometric variability or sharp transitions, the improvement from ACT may be less significant.

Second, although our visualizations and ablation results support the interpretation that adaptive coordinate transformations simplify the effective operator seen by subsequent latent updates, this claim is currently supported mainly by empirical evidence rather than by a formal theoretical characterization. A more rigorous understanding of when and why learned coordinate realignment improves operator learning remains an important direction for future work.

Conclusion

In this work, we introduced ACT as a lightweight and plug-and-play mechanism for geometric realignment in neural operator learning. Across multiple backbones and PDE benchmarks, the results show that adaptive coordinate transformations can substantially improve accuracy, especially when transport, mixing, or deformation is a dominant source of complexity. The qualitative analyses and ablations further suggest that ACT helps redistribute part of the geometric burden from the operator updates into the learned coordinate system.

Several directions remain open for future work. In particular, it is important to better understand the theoretical conditions under which learned coordinate realignment simplifies operator learning, and to extend the current framework beyond the present structured-grid setting to more general boundary conditions, complex geometries, and irregular meshes.

More broadly, our results suggest that coordinate choice can fundamentally shape the difficulty of operator learning, and the geometric realignment achieved by ACT offers a promising way to reduce this complexity.

Additional Implementation Details

This appendix records implementation details that complement, rather than repeat, the ACT formulation in § 3.3. In the current implementation, the backbone is instantiated as a 2D Fourier neural operator. The input is lifted by two pointwise convolutions into a hidden representation, processed by a stack of Fourier blocks, and then projected back to the output channels by another pointwise head. Each Fourier block consists of a spectral convolution, a skip connection, optional normalization, and an optional channel-mixing MLP. The implementation also supports gradient checkpointing on the Fourier blocks and an ablation mode in which ACT is disabled by replacing it with a pass-through module.

At the implementation level, each ACT block first applies a 1×11 \times1 projection to produce head-wise value features, then folds the head dimension into the batch dimension so that all heads are processed in parallel. If x(l)x^{(l)} denotes the hidden feature map at layer ll, the packed head features can be written as

V(l)=reshape⁡(Wvx(l))∈R(BM)×Ch×H×W.(18)V^{(l)} = \operatorname{reshape}(W_vx^{(l)}) \in\mathbb{R}^{(BM)\times C_h \times H \times W}. \tag*{(18)}

When coordinate conditioning is enabled, the cached identity-grid channels are concatenated to these head-wise value features before offset prediction. In the default branched setting, the displacement tower combines parallel 1×11 \times1, 3×33 \times3, and dilated 3×33 \times3 convolutions before a final 1×11 \times1 fusion layer outputs the offsets. The resulting displacement is bounded by

Δ(l)=αtanh⁡(fdisp([V(l),Xid])),(19)\Delta^{(l)} = \alpha\tanh\left(f_{\mathrm{disp}}\left([V^{(l)}, X_{\mathrm{id}}]\right)\right), \tag*{(19)}

where α\alpha is the prescribed maximum displacement. The last fusion layer is initialized to zero, so each ACT block starts exactly from the identity transformation. The code also includes a non-branched alternative in which the displacement tower is replaced by a two-layer 3×33 \times3 stack with GroupNorm and GELU.

The reference coordinates are cached as a normalized identity grid on [−1,1]2[-1,1]^2 and reused across calls as long as the spatial resolution and device do not change. The maximum displacement can either be specified manually or derived from the spatial resolution as the normalized size of one grid cell.

Resampling is carried out with bilinear grid_sample on normalized coordinates. If G(l)=Xid+Δ(l)G^{(l)} = X_{\mathrm{id}} + \Delta^{(l)} denotes the deformed sampling grid, the code applies a periodic fold only where coordinates are strictly outside the valid range,

G~(l)={G(l),if G(l)∈[−1,1]2,remainder⁡(G(l)+1,2)−1,otherwise.(20)\widetilde{G}^{(l)} = \begin{cases} G^{(l)}, & \text{if } G^{(l)} \in[-1,1]^2,\\ \operatorname{remainder}(G^{(l)}+1,2)-1, & \text{otherwise}. \end{cases} \tag*{(20)}

Hence coordinates already equal to ±1\pm1 remain unchanged, so zero predicted displacement reproduces the exact identity sampling pattern. The actual sampling call uses padding_mode=border and align_corners=True.

After resampling, the transformed head features are concatenated and projected back to the hidden width, then merged with a residual shortcut,

h(l+1)={Norm⁡(WoV^(l)+Wresx(l)),for non-final ACT blocks,WoV^(l)+Wresx(l),for the final ACT block.(21)h^{(l+1)} = \begin{cases} \operatorname{Norm}(W_o\widehat{V}^{(l)} + W_{\mathrm{res}}x^{(l)}), & \text{for non-final ACT blocks},\\ W_o\widehat{V}^{(l)} + W_{\mathrm{res}}x^{(l)}, & \text{for the final ACT block}. \end{cases} \tag*{(21)}

where V^(l)\widehat{V}^{(l)} denotes the concatenated sampled head features. This matches the implementation flag that skips output normalization in the last ACT block. A further training-related detail is that displacement tensors are not returned in training mode: each ACT block outputs (tensor, None) while self.training is true, and only exposes the displacement fields in evaluation mode. This design reduces the need to retain layer-wise displacement tensors during training.

Dataset Ilustrations

Table 3 provides a comprehensive overview of the datasets utilized in our experiments. We evaluate our models across a diverse suite of physical systems, encompassing both two-dimensional and three-dimensional Cartesian coordinates. The table details the fundamental properties of each dataset, including the spatial resolution and the temporal length of each trajectory (n_steps). Furthermore, to ensure rigorous and reproducible benchmarking, we explicitly outline the data partitioning strategy, providing the exact number of trajectories allocated for the training, validation, and test splits.

Table 3: Detailed description of the datasets, including coordinate system (CS), spatial resolution, number of time-steps per trajectory (n_steps), and the number of trajectories in the training, validation, and test sets.

Navier-Stokes Equation: Our benchmark includes the two-dimensional Navier-Stokes (NS) equations governing a viscous and incompressible fluid. The dynamics are expressed via the vorticity formulation on a unit torus domain:

∂tw(x,t)+u(x,t)⋅∇w(x,t)=νΔw(x,t)+f(x),x∈(0,1)2,t∈(0,T]∇⋅u(x,t)=0,x∈(0,1)2,t∈[0,T]w(x,0)=w0(x),x∈(0,1)2(22)\begin{aligned} \partial_t w(x,t) + u(x,t) \cdot\nabla w(x,t) &= \nu\Delta w(x,t) + f(x), && x \in(0,1)^2, t \in(0,T] \\ \nabla\cdot u(x,t) &= 0, && x \in(0,1)^2, t \in[0,T] \\ w(x,0) &= w_0(x), && x \in(0,1)^2 \tag*{(22)} \end{aligned}

In these governing equations, u∈C([0,T];Hperr((0,1)2;R2))u \in C([0,T];H_{\mathrm{per}}^r((0,1)^2;\mathbb{R}^2)) (with r>0r > 0) defines the fluid velocity, while w=∇×uw = \nabla\times u denotes the scalar vorticity field. The initial condition is given by w0∈Lper2((0,1)2;R)w_0 \in L_{\mathrm{per}}^2((0,1)^2;\mathbb{R}), ν∈R+\nu\in\mathbb{R}_+ controls the kinematic viscosity, and f∈Lper2((0,1)2;R)f \in L_{\mathrm{per}}^2((0,1)^2;\mathbb{R}) acts as a time-independent external forcing term. We aim to approximate the solution operator G†:ω(⋅,t0:tk)→ω(⋅,tk+1)G^\dagger: \omega(\cdot,t_0:t_k) \to\omega(\cdot,t_{k+1}).

We evaluate the models under a representative viscosity regime with ν=10−4\nu= 10^{-4}. For all experiments, both training and testing data are defined on a fixed spatial grid of resolution 128×128128 \times128.

2D Diffusion-Reaction Equation : Our study also employs the two-dimensional diffusion-reaction datasets provided by PDEBench [27]. This benchmark broadens the physical modeling to a 2D spatial domain and involves two non-linearly coupled quantities: the activator u=u(t,x,y)u = u(t,x,y) and the inhibitor v=v(t,x,y)v = v(t,x,y). The system’s dynamics are governed by the following partial differential equations:

∂tu=Du∂xxu+Du∂yyu+Ru,∂tv=Dv∂xxv+Dv∂yyv+Rv(23)\partial_t u = D_u \partial_{xx}u + D_u \partial_{yy}u + R_u,\qquad\partial_t v = D_v \partial_{xx}v + D_v \partial_{yy}v + R_v \tag*{(23)}

Here, DuD_u and DvD_v represent the constant diffusion coefficients associated with the activator and inhibitor, respectively. Meanwhile, Ru=Ru(u,v)R_u = R_u(u,v) and Rv=Rv(u,v)R_v = R_v(u,v) correspond to their localized reaction terms. The spatio-temporal domain is bounded by x,y∈(−1,1)x,y \in(-1,1) and spans a time interval of t∈(0,5]t \in(0,5]. Such coupled systems are widely recognized for their utility in simulating biological pattern generation. Specifically, the reaction mechanisms are modeled after the classic Fitzhugh-Nagumo kinetics [14], which are formulated as:

Ru(u,v)=u−u3−k−v,(24)R_u(u,v) = u - u^3 - k - v, \tag*{(24)}
Rv(u,v)=u−v,(25)R_v(u,v) = u - v, \tag*{(25)}

For these experiments, the scalar parameter is fixed at k=5×10−3k = 5 \times10^{-3}, and the respective diffusion rates are specified as Du=1×10−3D_u = 1 \times10^{-3} and Dv=5×10−3D_v = 5 \times10^{-3}. Initial states are synthesized by sampling from a standard normal distribution, such that u(0,x,y)∼N(0,1.0)u(0,x,y) \sim\mathcal{N}(0,1.0) across the spatial grid. While the native PDEBench repository supplies high-resolution data with grid dimensions of Nx=512N_x = 512, Ny=512N_y = 512, and Nt=501N_t = 501, we adopt the spatially and temporally downsampled variant (Nx=128,Ny=128,Nt=101)(N_x = 128, N_y = 128, N_t = 101) to facilitate efficient neural network training. To generate the ground truth, the spatial derivatives were discretized via the finite volume method [22], while temporal progression was handled by the standard fourth-order Runge-Kutta (RK4) integrator available in the scipy library [28]. Overall, the core objective is to learn the solution operator G†G^{\dagger} that maps the coupled historical states of the activator and inhibitor up to an initial time window to their subsequent evolution up to the next timestep. Formally, this operator we aim to learn is defined as G†:C([t0:tk];Hr((−1,1)2;R2))→C(tk+1;Hr((−1,1)2;R2))G^{\dagger}: C([t_0:t_k]; H^r((-1,1)^2;\mathbb{R}^2)) \to C(t_{k+1}; H^r((-1,1)^2;\mathbb{R}^2)).

Shallow Water Equations: For the shallow-water experiments, we adopt the benchmark configurations provided by PDEBench [27]. Formulated as a depth-averaged simplification of the compressible Navier-Stokes equations, the shallow-water equations (SWE) serve as a fundamental mathematical model for simulating free-surface fluid dynamics. Mathematically, this 2D hyperbolic PDE system is naturally represented in a vectorized conservation form:

∂th+∇⋅(hu)=0,(26)\partial_t h + \nabla\cdot(h\mathbf{u}) = 0, \tag*{(26)}
∂t(hu)+∇⋅(hu⊗u+12grh2I)=−grh∇b,(27)\partial_t(h\mathbf{u}) + \nabla\cdot\left(h\mathbf{u} \otimes\mathbf{u} + \frac{1}{2}g_r h^2\mathbf{I}\right) = -g_r h\nabla b, \tag*{(27)}

In this notation, u=(u,v)\mathbf{u}=(u,v) captures the two-dimensional velocity vector, hh stands for the localized water column depth, and bb denotes the underlying bottom topography (bathymetry). Consequently, the product huh\mathbf{u} signifies the momentum vector, while grg_r indicates the constant of gravitational acceleration. A critical property of this formulation is its strict preservation of mass and momentum conservation, even in the presence of discontinuous shock waves. This trait makes it highly suitable for generating complex, non-linear physical datasets that mirror real-world scenarios such as flash floods and tsunamis. By decomposing the vectors along the spatial axes, the governing dynamics can be explicitly formulated as:

∂th+∂xhu+∂yhv=0,(28)\partial_t h + \partial_x hu + \partial_y hv = 0, \tag*{(28)}
∂thu+∂x(u2h+12grh2)+∂yuvh=−grh∂xb,(29)\partial_t hu + \partial_x\left(u^2h+\frac{1}{2}g_rh^2\right) + \partial_y uvh = -g_rh\partial_xb, \tag*{(29)}
∂thv+∂y(v2h+12grh2)+∂xuvh=−grh∂yb,(30)\partial_t hv + \partial_y\left(v^2h+\frac{1}{2}g_rh^2\right) + \partial_xuvh = -g_rh\partial_yb, \tag*{(30)}

Regarding the specific experimental setup, the PDEBench data collection models a two-dimensional radial dam-break phenomenon. The simulation operates within a bounded square region Ω=[−2.5,2.5]2\Omega=[-2.5,2.5]^2. The initial fluid depth is structured as a centralized cylindrical column, governed by the following piecewise function:

h(t=0,x,y)={2.0,for r<x2+y21.0,for r≥x2+y2(31)h(t=0,x,y)= \begin{cases} 2.0, & \text{for } r < \sqrt{x^2+y^2}\\ 1.0, & \text{for } r \ge\sqrt{x^2+y^2} \end{cases} \tag*{(31)}

Here, the initial column radius rr is drawn stochastically from a uniform continuous distribution U(0.3,0.7)\mathcal{U}(0.3,0.7). The reference trajectories for these complex fluid interactions were computed via the robust finite volume schemes implemented within the PyClaw software library [13]. Specifically, the primary objective is to approximate the solution operator G†G^{\dagger} that maps the historical states of the water depth up to an initial time window to its future dynamics up to the next timestep, the operator we aim to learn as G†:C([t0:tk];Hr(Ω;R))→C(tk+1;Hr(Ω;R))G^{\dagger}: C([t_0:t_k];H^r(\Omega;\mathbb{R})) \to C(t_{k+1};H^r(\Omega;\mathbb{R})), tracking the temporal transition h∣Ω×[t0:tk]→h∣Ω,tk+1h|_{\Omega\times[t_0:t_k]} \to h|_{\Omega,t_{k+1}}, where Ω\Omega denotes the spatial domain.

Active Matter: We incorporate active matter datasets derived from the works of Ohana et al. [23]; Maddu et al. [20]. Fundamentally, active matter comprises autonomous entities—ranging from synthetic micro-swimmers to biological macromolecules—capable of transducing chemical energy into mechanical motion, consequently exerting local active stresses. Within the bulk, these generated forces propagate via mechanisms like hydrodynamic interactions, structural cross-linking, or direct steric hindrances, resulting in intricately complex spatiotemporal behaviors. The specific configurations adopted in our benchmark model a suspension of active particles within a highly viscous medium. This environment induces orientation-coupled viscosity and pronounced long-range interactions, thereby presenting a formidable challenge for operator learning architectures and data-driven surrogate modeling.

The foundational data generation simulates NN elongated active particles (with length ℓ\ell and width bb, satisfying ℓ/b≫1\ell/b \gg1) dispersed throughout a Stokes flow regime of volume VV. As the system approaches the macroscopic limit, the suspension dynamics are effectively captured by continuum kinetic frameworks tracking the probability density function Ψ(x,p,t)\Psi(\mathbf{x},\mathbf{p},t) [6, 8]. The temporal evolution of Ψ\Psi, which inherently conserves the total particle count, is dictated by the Smoluchowski equation:

∂Ψ∂t+∇x⋅(x˙Ψ)+∇p⋅(p˙Ψ)=0,(32)\frac{\partial\Psi}{\partial t}+\nabla_{\mathbf{x}}\cdot(\dot{\mathbf{x}}\Psi)+\nabla_{\mathbf{p}}\cdot(\dot{\mathbf{p}}\Psi)=0, \tag*{(32)}

Here, the variables x˙\dot{\mathbf{x}} and p˙\dot{\mathbf{p}} represent the translational and rotational fluxes, respectively, dictated by individual particle kinematics within the ambient velocity field u(x,t)\mathbf{u}(\mathbf{x},t). Integrating the distribution function over the orientational space (⟨f⟩=∫∣p∣=1fΨ dp\langle f\rangle=\int_{|\mathbf{p}|=1}f\Psi\,\mathrm{d}\mathbf{p}) allows us to extract macroscopic moments: the scalar concentration c=⟨1⟩c=\langle1\rangle, the polarity vector n=⟨p⟩/c\mathbf{n}=\langle\mathbf{p}\rangle/c, and the nematic order tensor Q=⟨pp⟩/c\mathbf{Q}=\langle\mathbf{p}\mathbf{p}\rangle/c. In the context of dense particle suspensions, these phase-space fluxes are formulated as:

x˙=u−dT∇xlog⁡Ψ;p˙=(I−pp)⋅(∇u+2ζD)⋅p−dR∇plog⁡Ψ.(33)\dot{\mathbf{x}}=\mathbf{u}-d_T\nabla_{\mathbf{x}}\log\Psi;\qquad\dot{\mathbf{p}}=(\mathbf{I}-\mathbf{p}\mathbf{p})\cdot(\nabla\mathbf{u}+2\zeta\mathbf{D})\cdot\mathbf{p}-d_R\nabla_{\mathbf{p}}\log\Psi. \tag*{(33)}

In these expressions, dTd_T and dRd_R denote the non-dimensional coefficients for translational and rotational diffusion, respectively. The parameter ζ\zeta quantifies the nematic alignment intensity driven by steric repulsion, while D=⟨pp⟩\mathbf{D}=\langle\mathbf{p}\mathbf{p}\rangle defines the second-order orientational moment. Furthermore, the fluid hydrodynamics are intricately coupled to the Smoluchowski dynamics via the incompressible Stokes equations:

−Δu+∇P=∇⋅Σ,∇⋅u=0,(34)-\Delta\mathbf{u}+\nabla P=\nabla\cdot\boldsymbol{\Sigma},\qquad\nabla\cdot\mathbf{u}=0, \tag*{(34)}
Σ=αD+βS:E−2ζβ(D⋅D−S:D).(35)\boldsymbol{\Sigma}=\alpha\mathbf{D}+\beta\mathbf{S}:\mathbf{E}-2\zeta\beta(\mathbf{D}\cdot\mathbf{D}-\mathbf{S}:\mathbf{D}). \tag*{(35)}

Within this fluid coupling, P(x,t)P(\mathbf{x},t) stands for the hydrodynamic pressure. The parameters α\alpha and β\beta scale the non-dimensional active dipole magnitude and the effective particle concentration, respectively. Additionally, E=[∇u+∇uT]/2\mathbf{E}=[\nabla\mathbf{u}+\nabla\mathbf{u}^{\mathsf{T}}]/2 calculates the symmetric strain-rate tensor, and S=⟨pppp⟩\mathbf{S}=\langle\mathbf{p}\mathbf{p}\mathbf{p}\mathbf{p}\rangle corresponds to the fourth-order orientational moment.

The coupled governing equations (32)–(34) were numerically resolved via a pseudo-spectral framework utilizing 256×256256\times256 spatial modes on a doubly periodic domain of size (0,10)2(0,10)^2, integrated temporally with a second-order semi-implicit backward differentiation scheme (SBDF2). Key dimensionless parameters were systematically modulated: the active contractility coefficient was uniformly sampled from α∈{−1,−2,−3,−4,−5}\alpha\in\{-1,-2,-3,-4,-5\}, the effective concentration parameter was fixed at β=0.8\beta=0.8, and the nematic alignment factor traversed ζ∈{1,3,5,7,9,11,13,15,17}\zeta\in\{1,3,5,7,9,11,13,15,17\}. The final curated dataset encapsulates an extensive suite of macroscopic physical fields: scalar concentration, 2D fluid velocity, orientational order tensor, and the symmetric strain-rate tensor. Formally, the target operator is defined as G†:C([t0:tk];Hperr((0,1)2;Rd))→C(tk+1;Hperr((0,1)2;Rd))G^{\dagger}:C([t_0:t_k];H_{\mathrm{per}}^r((0,1)^2;\mathbb{R}^d))\to C(t_k+1;H_{\mathrm{per}}^r((0,1)^2;\mathbb{R}^d)), tracking the temporal transition of the combined macroscopic state vector ϕ:ϕ(0,1)2×[t0:tk]↦ϕ(0,1)2,tk+1\phi:\phi_{(0,1)^2\times[t_0:t_k]}\mapsto\phi_{(0,1)^2,t_{k+1}}.

Turbulent Radiative Layer (2D): The benchmark additionally incorporates the 2D turbulent mixing layer dataset from The Well [23], based on the foundational simulations by [7]. This dataset models the interfacial dynamics between cold, dense gas and ambient hot plasma, where the survival of cold structures is dictated by the competition between turbulent mixing and rapid radiative cooling [9, 1]. The primary objective is to learn the forward solution operator G†:C([t0:tk];Hr(Ω;R4))→C(tk+1;Hr(Ω;R4))G^{\dagger}:C([t_0:t_k];H^r(\Omega;\mathbb{R}^4))\to C(t_{k+1};H^r(\Omega;\mathbb{R}^4)), which tracks the temporal transition of the coupled fluid fields across the 2D spatial domain Ω\Omega: (ρ,P,v⃗)∣Ω×[t0:tk]↦(ρ,P,v⃗)∣Ω,tk+1(\rho,P,\vec{v})|_{\Omega\times[t_0:t_k]}\mapsto(\rho,P,\vec{v})|_{\Omega,t_{k+1}}.

Mathematically, the system governs the hydrodynamic conservation laws augmented with a temperature-dependent energy sink, parameterized by a characteristic cooling time tcoolt_{\mathrm{cool}}:

∂ρ∂t+∇⋅(ρv⃗)=0(36)\frac{\partial\rho}{\partial t}+\nabla\cdot(\rho\vec{v})=0 \tag*{(36)}
∂ρv⃗∂t+∇⋅(ρv⃗v⃗+P)=0(37)\frac{\partial\rho\vec{v}}{\partial t}+\nabla\cdot(\rho\vec{v}\vec{v}+P)=0 \tag*{(37)}
∂E∂t+∇⋅((E+P)v⃗)=−Etcool(38)\frac{\partial E}{\partial t}+\nabla\cdot((E+P)\vec{v})=-\frac{E}{t_{\mathrm{cool}}} \tag*{(38)}
E=Pγ−1whereγ=53(39)E=\frac{P}{\gamma-1}\qquad\text{where}\qquad\gamma=\frac{5}{3} \tag*{(39)}

A fundamental property of this system is the scaling relationship between the volume-integrated cooling rate, the net mass accretion flux, the relative velocity, and the cooling timescale, expressed as

E˙cool∝M˙∝vrel3/4tcool−1/4.\dot{E}_{\mathrm{cool}} \propto\dot{M} \propto v_{\mathrm{rel}}^{3/4}t_{\mathrm{cool}}^{-1/4}.

To systematically sample this dynamic landscape, the underlying high-fidelity simulations were generated by sweeping the cooling timescale across

tcool∈{0.03,0.06,0.1,0.18,0.32,0.56,1.00,1.78,3.16}.t_{\mathrm{cool}} \in\{0.03, 0.06, 0.1, 0.18, 0.32, 0.56, 1.00, 1.78, 3.16\}.

The finalized dataset captures the temporal evolution of the 2D density, pressure, and velocity fields, efficiently circumventing the substantial computational overhead required for conventional numerical integration.

Magnetohydrodynamic (MHD) Turbulence: We integrate the magnetohydrodynamic (MHD) turbulence dataset directly from The Well repository [23]. Turbulence within MHD systems acts as a cornerstone for numerous astrophysical phenomena, dictating processes from interstellar medium (ISM) evolution to large-scale galactic formation. Specifically, this dataset features isothermal, non-self-gravitating MHD flows that accurately emulate the physical environment of the diffuse ISM. The primary objective is to approximate the solution operator G†G^{\dagger} that maps the temporal history of the coupled magnetohydrodynamic fields—comprising the scalar density, the 3D fluid velocity vector, and the 3D magnetic field vector—up to an initial time window to their subsequent turbulent evolution up to a final time.

The governing ideal MHD equations, resolved via a third-order hybrid essentially non-oscillatory (ENO) scheme [3], formulate the conservation laws for mass, momentum, and magnetic induction as follows:

∂ρ∂t+∇⋅(ρv)=0,(40)\frac{\partial\rho}{\partial t}+\nabla\cdot(\rho\mathbf{v})=0, \tag*{(40)}
∂ρv∂t+∇⋅[ρvv+(p+B28π)I−14πBB]=f,(41)\frac{\partial\rho\mathbf{v}}{\partial t}+\nabla\cdot\left[\rho\mathbf{v}\mathbf{v}+\left(p+\frac{B^2}{8\pi}\right)\mathbf{I}-\frac{1}{4\pi}\mathbf{B}\mathbf{B}\right]=\mathbf{f}, \tag*{(41)}
∂B∂t−∇×(v×B)=0.(42)\frac{\partial\mathbf{B}}{\partial t}-\nabla\times(\mathbf{v}\times\mathbf{B})=0. \tag*{(42)}

In this system, ρ\rho signifies the fluid density, v\mathbf{v} defines the local velocity vector, B\mathbf{B} represents the magnetic field, pp denotes the isotropic gas pressure (p=cs2ρp=c_s^2\rho with isothermal sound speed csc_s), and I\mathbf{I} acts as the identity tensor. Continuous turbulence is sustained by the external source term f\mathbf{f}, which operates as a stochastic solenoidal driving force peaking at the wavenumber k≈2.5k\approx2.5.

To ensure a robust exploration of the chaotic parameter space, the dataset provides a diverse spectrum of turbulent flows systematically modulated across two key dimensionless parameters: the sonic Mach number (Ms≡∣v∣/cs∈{0.5,0.7,1.5,2.0,7.0}\mathcal{M}_s\equiv|\mathbf{v}|/c_s\in\{0.5,0.7,1.5,2.0,7.0\}) and the Alfvénic Mach number (MA≡∣v∣/⟨vA⟩∈{0.7,2.0}\mathcal{M}_A\equiv|\mathbf{v}|/\langle v_A\rangle\in\{0.7,2.0\}). These configurations encompass both magnetically-dominated sub-Alfvénic regimes and super-Alfvénic environments. The generated snapshots, curated at a 64364^3 spatial resolution to balance fidelity with computational tractability, provide a highly comprehensive multi-modal dataset consisting of scalar mass density, vector fluid velocity, and vector magnetic fields. Formally, this operator is defined as G†:C([t0:tk];Hperr(Ω;R7))→C(tk+1;Hperr(Ω;R7))G^{\dagger}:C([t_0:t_k];H_{\mathrm{per}}^r(\Omega;\mathbb{R}^7))\to C(t_{k+1};H_{\mathrm{per}}^r(\Omega;\mathbb{R}^7)), where Ω\Omega denotes the 3D periodic spatial domain, tracking the temporal transition (ρ,v,B)∣Ω×[t0,tk]↦(ρ,v,B)∣Ω,tk+1(\rho,\mathbf{v},\mathbf{B})|_{\Omega\times[t_0,t_k]}\mapsto(\rho,\mathbf{v},\mathbf{B})|_{\Omega,t_{k+1}}, where Ω\Omega denotes the 3D periodic spatial domain.

Model Structure and Hyperparameters

ACT-CNextU

As detailed in Table 4, the proposed Adaptive Coordinate Transformation (ACT) enhances the ConvNeXt-UNet [18] architecture by systematically integrating a MultiHeadACTBlock after each encoder and decoder stage. The vanilla ConvNeXt-UNet serves as a robust multi-scale backbone, comprising 18,571,562 trainable parameters. Enabling the ACT mechanism across all 8 hierarchical stages introduces an additional 997,878 parameters. This corresponds to an overhead of approximately 5.37% relative to the vanilla model, yielding a total of 19,569,440 parameters for the enhanced UNetConvNextACT. This marginal increase confirms that the performance improvements achieved by our method are fundamentally driven by the structured, coordinate-aware spatial transformations introduced by ACT, rather than a mere scaling of the model capacity.

ComponentParamsNotes
Input / Stem
Initial convolutions: Conv2d ×\times 23,8243,066 + 758
Vanilla ConvNeXt-UNet Architecture (Subtotals)
Encoder (4 stages): Downsample + 2×\timesConvNeXt blocks3,674,156feature encoding
Bottleneck: 1×\timesConvNeXt block3,651,648latent representation
Decoder (4 stages): Conv2d + Upsample + 2×\timesConvNeXt blocks11,241,934feature decoding
Vanilla Backbone subtotal18,567,738multi-scale UNet
MultiHeadACTBlock (ACT, distributed across 8 stages)apply ACT
value_proj Conv2d485,888channel mixing
disp branches (ModuleDict): 1x1, 3x3, dilated 3x322,0808×\times(152+1304+1304)
disp_fuse: GELU + Conv2d(24→\to2)4008×\times50
sampled with augmented grid–
out_proj Conv2d485,730project features to output
GroupNorm3,780output normalization
ACT modules subtotal997,8788 ACT blocks in total
Total trainable params (apply ACT)19,569,440UNetConvNextACT
Total trainable params (vanilla UNet)18,571,562use_ACT=False
Optional ACT overhead+997,878+5.37%

Table 4. Detailed parameter breakdown for the ConvNeXt-UNet enhanced with the ACT module (ACT-CNextU). The ACT mechanism is composed of MultiHeadACTBlocks distributed across the encoder and decoder stages. Disabling them recovers the vanilla ConvNeXt-UNet.

We further emphasize that the integration of ACT is seamlessly adapted to the hierarchical nature of the U-Net architecture. Throughout all experimental evaluations, the underlying multi-scale encoder-bottleneck-decoder structure remains strictly fixed. While the offset generation branches within the ACT modules share a highly efficient, constant parameter footprint across stages, the linear projection layers naturally scale with the corresponding feature map dimensions and channel counts. Consequently, the computational cost introduced by the ACT modules is optimally distributed across the network’s resolutions. This demonstrates that ACT is a highly efficient, plug-and-play enhancement that can be gracefully applied to modern hierarchical vision backbones without disrupting their original design principles.

ACT-Translover

As detailed in Table 5, the proposed Adaptive Coordinate Transformation (ACT) augments the Transolver backbone by integrating a MultiHeadACTBlock after each standard TransolverBlock. Because the vanilla Transolver is an inherently lightweight architecture, comprising only 188,198 trainable parameters, enabling the ACT mechanism introduces an additional 144,104 parameters. Although this corresponds to a 76.57% relative parameter overhead, the absolute increase remains highly tractable, resulting in a total of just 332,302 parameters for the enhanced model (TransolverPlus). This demonstrates that the performance improvements achieved by our method are fundamentally driven by the structured, coordinate-aware feature transformations introduced by ACT, rather than relying on massive parameter scaling or heavy model capacities.

ComponentParamsNotes
Input / Stem
Global Parameter (Transolver root)64global token / state
TransolverMLP: Linear + GELU + Linear9,6641,408 + 8,256 (feature lifting)
Backbone stage (repeated 4 times)
Step kk (k=1…4k = 1 \ldots4):–
TransolverBlock (single block details)
LayerNorm128pre-normalization
PhysicsAttention1DEidetic9,129physics-attention
LayerNorm128pre-normalization
TransolverMLP: Linear + GELU + Linear33,08816,640 + 16,448
TransolverBlock subtotal42,473per block
MultiHeadACTBlock (when apply ACT)apply ACT, if disabled, this becomes vanilla Transolver
value_proj Conv2d16,640256→64256 \to64
disp branches (ModuleDict): 1x1, 3x3, dilated 3x32,760channel mixing
disp_fuse: GELU + Conv2d(24→224 \to2)50offset prediction
sampled with augmented grid–
out_proj Conv2d16,448project features to output channels
GroupNorm128output normalization
ACTBlock subtotal36,026per block
Backbone total (with ACT): 4×4 \times(TransolverBlock + ACTBlock)313,9964×(42,473+36,026)4 \times(42,473 + 36,026)
Backbone total (vanilla Transolver): 4×4 \timesTransolverBlock169,892ACT disabled
Output Head for ACT-Transolver
Projection: Conv2d + GELU + Conv2d8,5788,320 + 258
Output Head for Vanilla Transolver
Projection: LayerNorm + Linear258
Total trainable params (apply ACT)332,302TransolverPlus
Total trainable params (vanilla Transolver)179,878use_ACT=False
Optional ACT overhead+144,104+76.57%

Table 5. Detailed parameter breakdown for Transolver enhanced with the ACT module(ACT-Transolver). MultiHeadACTBlock acts as the ACT mechanism; disabling it recovers the vanilla Transolver architecture.

We further emphasize that, throughout all experimental evaluations, our Transolver-based models strictly adhere to the unified architectural configuration presented in Table 5. When deploying the model across different datasets, the core backbone architecture remains completely fixed. The only source of variation in the overall parameter count arises from the input lifting stem and the output projection head, whose dimensions depend solely on the dataset-specific input and output channels. Consequently, the parameter overhead introduced by the ACT blocks is essentially invariant across different datasets. Therefore, reporting a single representative parameter breakdown is sufficient to accurately characterize the architectural modifications and the additional computational cost introduced by ACT under all experimental settings.

ACT-FNO

As summarized in Table 6, the proposed Adaptive Coordinate Transformation (ACT) augments the FNO backbone by inserting a lightweight MultiHeadACTBlock after each backbone stage, incurring only a marginal increase in trainable parameters. Concretely, enabling the ACT block introduces an additional 45,032 parameters, which corresponds to approximately 0.27% overhead relative to the vanilla FNO backbone. This clearly indicates that the performance improvements achieved by ACT are not attributed to increased model capacity, but rather to more effective and structured coordinate-aware feature transformations.

ComponentParamsNotes
Input / Stem
Lifting: Conv1x1(Cin→128C_{in}\rightarrow128) + GELU + Conv1x1(128→64128\rightarrow64)9,408feature lifting
Backbone stage (repeated 4 times, alternating)
Step kk (k=1…4k = 1 \ldots4):–
FNOBlock2d (single block details)
SpectralConv2d(modes 32, channel 64)4,194,304dominant spectral weights
Conv2d (skip / linear)4,096channel mixing
GroupNorm ×2256128 + 128
MLP: Conv1x1(64→3264\rightarrow32) + GELU + Conv1x1(32→6432\rightarrow64)4,1922,080 + 2,112
SoftGating64per-channel gate
FNOBlock2d subtotal4,202,912per block
MultiHeadACTBlock (when apply ACT)apply ACT, if disabled, this becomes vanilla FNO
value_proj Conv1x14,16064→6464\rightarrow64
disp branches (ModuleDict): 1x1, 3x3, dilated 3x32,760channel mixing
disp_fuse: GELU + Conv1x1(24→224\rightarrow2)50offset prediction
sampled with arguemented grid–
out_proj Conv1x14,160project features to output channels
GroupNorm128output normalization
ACTBlock subtotal11,258per block
Backbone total (with ACT): 4×4\times(FNOBlock2d + ACTBlock)16,856,6804×(4,202,912+11,258)4 \times(4,202,912 + 11,258)
Backbone total (vanilla FNO): 4×4\timesFNOBlock2d16,811,648ACT disabled
Head
Projection: Conv1x1(64→12864\rightarrow128) + GELU + Conv1x1(128→Cout128\rightarrow C_{out})8,5788,320 + 258
Total trainable params (apply ACT)16,874,666ACT-FNO
Total trainable params (vanilla FNO)16,829,634use_ACT=False
Optional ACT overhead+45,032+0.27% only

Table 6. Layer-by-layer parameter breakdown of the proposed ACT-FNO architecture. The MultiHeadACTBlock acts as a lightweight, plug-and-play enhancement; removing it completely recovers the standard vanilla FNO baseline.

We further emphasize that, throughout all experimental evaluations, our FNO-based models strictly adhere to the unified architectural configuration presented in Table 6. When deploying the model across different datasets, the backbone architecture remains fixed, and the only source of variation in parameter count arises from the input lifting layer and the output projection head, whose dimensions depend solely on the dataset-specific input and output channels. Consequently, the relative parameter overhead introduced by the ACT block is essentially invariant across different datasets. Therefore, reporting a single representative parameter breakdown is sufficient to accurately characterize the additional computational cost introduced by ACT under all experimental settings.

Common Experiment Settings

To ensure a rigorous and fair evaluation, all compared models are trained and tested under a strictly unified experimental pipeline. The comprehensive list of common hyperparameters and configuration details is summarized in Table 7. Specifically, the data formulation is fixed to utilize 4 historical steps to predict the subsequent 1 step, with consistent Z-score normalization and temporal strides. Furthermore, the optimization dynamics—including the AdamW optimizer, learning rate schedules, batch size, and total training epochs—are held strictly constant across all baselines and our proposed models. Crucially, by keeping these training and validation protocols identical, we isolate the network architecture as the sole independent variable. This guarantees that any observed performance improvements are strictly attributable to the architectural enhancements (such as the integration of the ACT module) rather than dataset-specific hyperparameter tuning or training tricks.

CategoryHyperparameterValue
ACT Blockmaximum displacement magnitude (Transolver&FNO)0.5
ACT Blockmaximum displacement magnitude (Unet)adaptive gird width (2Resolution of Input)\left(\frac{2}{\text{Resolution of Input}}\right)
DataInput steps4
DataOutput steps1
DataNormalizationZ-score normalization
DataTemporal stridestride = 1
BatchingBatch size32
OptimizationOptimizerAdamW
OptimizationLearning rate5×10−45 \times10^{-4}
OptimizationWeight decay1×10−41 \times10^{-4}
LR scheduleSchedulerStepLR
LR scheduleStep size / decay factor for 2d data50 / 0.5
LR scheduleStep size / decay factor for 3d data25 / 0.5
TrainingEpochs for 2d data except Active_Matter300
TrainingEpochs for Active_Matter200
TrainingEpochs 3d data100
TrainingLoss functionMSE
TrainingGradient clipping1.0
TrainingFormatterchannels-first
TrainingCheckpoint frequencyevery 10 epochs
ValidationValidation frequencyevery 20 epochs
ValidationShort validation length30 batches

Table 7. Common experimental settings shared by all compared models. Only the model architecture is changed across experiments, ensuring a fair comparison.

Computing Configuration

As shown in Table 8, all models and physical simulations were executed on a high-performance computing (HPC) cluster running a Linux-based operating system. The compute nodes are equipped with modern multi-core server-grade processors and large-memory configurations to support data-intensive workloads.

ComponentSpecification
Operating SystemLinux-based HPC environment
CPUMulti-core server-grade processors (node-dependent)
GPU4× NVIDIA A100 (80 GB VRAM or 40 GB VRAM)
System MemoryHigh-memory configuration (node-dependent)

Table 8. Hardware configuration of the GPU-enabled compute nodes used for model training.

To accelerate neural network training and large-scale data generation, GPU-enabled nodes were utilized. Each node is equipped with four NVIDIA A100 Tensor Core GPUs, each providing 80 GB or 40 GB of VRAM, enabling efficient distributed and memory-intensive computations.

To evaluate the computational efficiency of our proposed method, Table 9 profiles the model parameters, GPU memory footprint, and single-forward inference latency across different baseline architectures on the shallow water dataset. The results demonstrate that the integration of the ACT module introduces merely a marginal increase in learnable parameters (e.g., adding only 0.04M parameters to FNO). While the dynamic coordinate reparameterization inherently incurs an expected overhead in memory consumption and forward inference time, the overall computational cost remains well within a practical range. Given the substantial improvements in predictive accuracy, this modest computational trade-off is highly justified.

ModelParams (M)Memory (MiB)Forward (s)
CNextU18.5711200.038
CNextU + ACT19.5742780.072
FNO16.8314100.029
FNO + ACT16.8718120.054
Transolver0.1836400.128
Transolver + ACT0.3346060.210

Table 9. GPU memory footprint and single-forward inference time on shallow water dataset. All experiments use a batch size of 32. Forward time is the median over 3 measured batches after 1 warmup batch.

Derivation of Transformed Operator Representations

This appendix expands the discussion in § 5.1 and shows explicitly how coordinate transformations modify the kernel representation of an integral operator. The notation follows Section 3: ψin\psi_{\mathrm{in}} and ψout\psi_{\mathrm{out}} denote forward coordinate transformations, while ϕin:=ψin−1\phi_{\mathrm{in}} := \psi_{\mathrm{in}}^{-1} and ϕout:=ψout−1\phi_{\mathrm{out}} := \psi_{\mathrm{out}}^{-1} denote the associated inverse maps. In the ACT instantiation of § 3.3, these abstract transformations are realized implicitly through the layer-wise blocks Tθ(l)T_{\theta}^{(l)} and their head-specific sampling maps ϕh(m)\phi_{h}^{(m)}.

Consider

F:u↦v,v(x)=∫ΩK(x,y)u(y) dy.(43)\mathcal{F}: u \mapsto v,\qquad v(x)=\int_{\Omega}K(x,y)u(y)\,\mathrm{d}y. \tag*{(43)}

Throughout this appendix, we assume that the coordinate transformations are bijective and sufficiently smooth so that the standard change-of-variables formula applies. The goal is not to change the underlying operator itself, but to rewrite the same mapping in different coordinates and make explicit how the kernel representation is modified.

Input-side transformation. Let ψin:Ω→Ω\psi_{\mathrm{in}}:\Omega\to\Omega be an input-side coordinate transformation, define transformed coordinates by ξ=ψin(y)\xi=\psi_{\mathrm{in}}(y), and let ϕin:=ψin−1\phi_{\mathrm{in}}:=\psi_{\mathrm{in}}^{-1} denote the associated inverse map. The transformed input is

uˉ(ξ)=u(ϕin(ξ)).(44)\bar{u}(\xi)=u\left(\phi_{\mathrm{in}}(\xi)\right). \tag*{(44)}

Equivalently, u(y)=uˉ(ψin(y))u(y)=\bar{u}(\psi_{\mathrm{in}}(y)), so the operator can be viewed as acting on the pulled-back input uˉ\bar{u} instead of the original field uu. Since only the source variable yy is re-parameterized, the evaluation coordinate xx remains unchanged. Applying the change of variables y=ϕin(ξ)y=\phi_{\mathrm{in}}(\xi), which leads to

dy=∣det⁡∇ϕin(ξ)∣ dξ,(45)\mathrm{d}y=\left|\det\nabla\phi_{\mathrm{in}}(\xi)\right|\,\mathrm{d}\xi, \tag*{(45)}

where ∇ϕin\nabla\phi_{\mathrm{in}} is the Jacobian matrix of ϕin\phi_{\mathrm{in}}, and det⁡\det denotes the determinant. Substituting this relation into the operator gives

v(x)=∫ΩK(x,y)u(y) dy=∫ΩK(x,ϕin(ξ))u(ϕin(ξ))∣det⁡∇ϕin(ξ)∣ dξ=∫ΩK(x,ϕin(ξ))uˉ(ξ)∣det⁡∇ϕin(ξ)∣ dξ=∫ΩK~in(x,ξ)uˉ(ξ) dξ,(46)\begin{aligned} v(x) &=\int_{\Omega}K(x,y)u(y)\,\mathrm{d}y\\ &=\int_{\Omega}K\left(x,\phi_{\mathrm{in}}(\xi)\right)u\left(\phi_{\mathrm{in}}(\xi)\right)\left|\det\nabla\phi_{\mathrm{in}}(\xi)\right|\,\mathrm{d}\xi\\ &=\int_{\Omega}K\left(x,\phi_{\mathrm{in}}(\xi)\right)\bar{u}(\xi)\left|\det\nabla\phi_{\mathrm{in}}(\xi)\right|\,\mathrm{d}\xi\\ &=\int_{\Omega}\widetilde{K}_{\mathrm{in}}(x,\xi)\bar{u}(\xi)\,\mathrm{d}\xi, \tag*{(46)} \end{aligned}

with the transformed kernel

K~in(x,ξ)=K(x,ϕin(ξ))∣det⁡∇ϕin(ξ)∣.(47)\widetilde{K}_{\mathrm{in}}(x,\xi)=K\left(x,\phi_{\mathrm{in}}(\xi)\right)\left|\det\nabla\phi_{\mathrm{in}}(\xi)\right|. \tag*{(47)}

Therefore, an input-side transformation induces a new operator representation whose kernel incorporates two effects at once: the source location yy is replaced by its expression in the transformed coordinates, and the Jacobian factor compensates for the change in volume element. This is why input-side transformations modify how information is aggregated over the source domain.

Output-side transformation. Let ψout:Ω→Ω\psi_{\mathrm{out}}:\Omega\to\Omega be an output-side coordinate transformation, define transformed coordinates by ξ=ψout(x)\xi=\psi_{\mathrm{out}}(x), and let ϕout:=ψout−1\phi_{\mathrm{out}}:=\psi_{\mathrm{out}}^{-1} denote the associated inverse map. The transformed output is

vˉ(ξ)=v(ϕout(ξ)).(48)\bar{v}(\xi)=v\left(\phi_{\mathrm{out}}(\xi)\right). \tag*{(48)}

Here the situation is different from the input-side case: the integration variable yy is untouched, so no Jacobian term appears. We simply evaluate the original output field at the physical point x=ϕout(ξ)x=\phi_{\mathrm{out}}(\xi) corresponding to the transformed coordinate ξ\xi. Substituting this relation into the operator gives

vˉ(ξ)=v(ϕout(ξ))=∫ΩK(ϕout(ξ),y)u(y) dy.(49)\bar{v}(\xi)=v\left(\phi_{\mathrm{out}}(\xi)\right)=\int_{\Omega}K\left(\phi_{\mathrm{out}}(\xi),y\right)u(y)\,\mathrm{d}y. \tag*{(49)}

This corresponds to the transformed kernel

K~out(ξ,y)=K(ϕout(ξ),y).(50)\widetilde{K}_{\mathrm{out}}(\xi,y)=K\left(\phi_{\mathrm{out}}(\xi),y\right). \tag*{(50)}

Hence, an output-side transformation changes the representation of the operator only through the query variable. In other words, it does not alter how source contributions are integrated, but only where the operator is sampled.

Summary. The two transformations act on complementary variables of the integral operator. Input-side transformations modify how information is aggregated from the source domain, while output-side transformations modify where the operator is evaluated. In both cases, the underlying mapping is unchanged, but the kernel representation seen by the learner is altered.

Supplementary Visualization Results

In this section, we provide comprehensive visualization results for the remaining benchmark datasets evaluated in our study. Specifically, we present qualitative comparisons between the vanilla baseline architectures (Transolver, FNO, and CNextU) and their ACT-augmented counterparts. The following figures illustrate the model predictions across various physical scenarios, including the scalar vorticity field ww in the Navier_Stokes dataset, the activator uu in the Diff_Reaction dataset, the fluid depth hh in the Shallow_Water dataset, the scalar concentration cc in the Active_Matter dataset, and the fluid density ρ\rho in the MHD_64 dataset. These supplementary visual comparisons further corroborate the enhanced spatial representation and fine-grained predictive accuracy introduced by the ACT module across diverse physical dynamics.

Target, predicted, and error visualizations for ACT-Transolver and VAN-Transolver on the Navier_Stokes dataset

Figure 4. The scalar vorticity field ww on the Navier_Stokes dataset comparing the vanilla Transolver and ACT-Transolver.

Target, predicted, and error visualizations for ACT-FNO and VAN-FNO on the Navier_Stokes dataset

Figure 5. The scalar vorticity field ww on the Navier_Stokes dataset comparing the vanilla FNO and ACT-FNO.

Scalar vorticity field comparison on the Navier\_Stokes dataset

Figure 6. The scalar vorticity field ww on the Navier_Stokes dataset comparing the vanilla CNextU and ACT-CNextU.

Activator comparison on the Diff\_Reaction dataset for Transolver

Figure 7. The activator uu on the Diff_Reaction dataset comparing the vanilla Transolver and ACT-Transolver.

Activator comparison on the Diff\_Reaction dataset for FNO

Figure 8. The activator uu on the Diff_Reaction dataset comparing the vanilla FNO and ACT-FNO.

Target, predicted, and absolute error visualizations for ACT-CNextU and VAN-CNextU on the Diff_Reaction dataset

Figure 9. The activator uu on the Diff_Reaction dataset comparing the vanilla CNextU and ACT-CNextU.

Target, predicted, and absolute error visualizations for ACT-Transolver and VAN-Transolver on the Shallow_Water dataset

Figure 10. The fluid depth hh on the Shallow_Water dataset comparing the vanilla Transolver and ACT-Transolver.

Target, predicted, and absolute error visualizations for ACT-FNO and VAN-FNO on the Shallow_Water dataset

Figure 11. The fluid depth hh on the Shallow_Water dataset comparing the vanilla FNO and ACT-FNO.

Comparison of fluid depth on the Shallow_Water dataset for vanilla CNextU and ACT-CNextU

Figure 12. The fluid depth hh on the Shallow_Water dataset comparing the vanilla CNextU and ACT-CNextU.

Comparison of scalar concentration on the Active_Matter dataset for vanilla Transolver and ACT-Transolver

Figure 13. The scalar concentration cc on the Active_Matter dataset comparing the vanilla Transolver and ACT-Transolver.

Comparison of scalar concentration on the Active_Matter dataset for vanilla FNO and ACT-FNO

Figure 14. The scalar concentration cc on the Active_Matter dataset comparing the vanilla FNO and ACT-FNO.

Comparison of target, predicted, and error scalar concentration fields for ACT-CNextU and VAN-CNextU

Figure 15. The scalar concentration cc on the Active_Matter dataset comparing the vanilla CNextU and ACT-CNextU.

Comparison of target, predicted, and error fluid density fields for ACT-Transolver and VAN-Transolver

Figure 16. The fluid density ρ\rho on the MHD_64 dataset comparing the vanilla Transolver and ACT-Transolver.

Comparison of target, predicted, and error fluid density fields for ACT-FNO and VAN-FNO

Figure 17. The fluid density ρ\rho on the MHD_64 dataset comparing the vanilla FNO and ACT-FNO.

The fluid density $\rho$ on the MHD\_64 dataset comparing the vanilla CNextU and ACT-CNextU.

Figure 18. The fluid density ρ\rho on the MHD_64 dataset comparing the vanilla CNextU and ACT-CNextU.

NeurIPS Paper Checklist

  1. Claims

Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope?

Answer: [Yes]

Justification: The abstract and introduction clearly state the main contributions and scope.

Guidelines:

  • The answer [N/A] means that the abstract and introduction do not include the claims made in the paper.

  • The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A [No] or [N/A] answer to this question will not be perceived well by the reviewers.

  • The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings.

  • It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper.

  1. Limitations

Question: Does the paper discuss the limitations of the work performed by the authors?

Answer: [Yes]

Justification: The paper discusses limitations in the Conclusion and Limitation section.

Guidelines:

  • The answer [N/A] means that the paper has no limitation while the answer [No] means that the paper has limitations, but those are not discussed in the paper.

  • The authors are encouraged to create a separate “Limitations” section in their paper.

  • The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be.

  • The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated.

  • The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be

used reliably to provide closed captions for online lectures because it fails to handle technical jargon.

  • The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size.

  • If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness.

  • While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren’t acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations.

Theory assumptions and proofs

Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof?

Answer: [N/A]

Justification: The paper does not present formal theorems or proofs; instead, it provides conceptual and analytical insights into how coordinate transformations affect operator representations.

Guidelines:

  • The answer [N/A] means that the paper does not include theoretical results.

  • All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced.

  • All assumptions should be clearly stated or referenced in the statement of any theorems.

  • The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition.

  • Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material.

  • Theorems and Lemmas that the proof relies upon should be properly referenced.

Experimental result reproducibility

Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)?

Answer: [Yes]

Justification: The paper provides detailed descriptions of model architectures, datasets, evaluation metrics, and training protocols sufficient to reproduce the results.

Guidelines:

  • The answer [N/A] means that the paper does not include experiments.

  • If the paper includes experiments, a [No] answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not.

  • If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable.

  • Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general, releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed.

  • While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example (a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. (b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. (c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). (d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results.

Open access to data and code

Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material?

Answer: [No]

Justification: While we plan to release the code upon publication, it is not included in the current submission to preserve anonymity. All implementation details are described to support reproducibility.

Guidelines:

  • The answer [N/A] means that paper does not include experiments requiring code.

  • Please see the NeurIPS code and data submission guidelines (https://neurips.cc/public/guides/CodeSubmissionPolicy) for more details.

  • While we encourage the release of code and data, we understand that this might not be possible, so [No] is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark).

  • The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://neurips.cc/public/guides/CodeSubmissionPolicy) for more details.

  • The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc.

  • The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why.

  • At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable).

  • Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted.

Experimental setting/details

Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer) necessary to understand the results?

Answer: [Yes]

Justification: Training details, datasets, model configurations, and evaluation protocols are clearly specified in the main text and appendix.

Guidelines:

  • The answer [N/A] means that the paper does not include experiments.

  • The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them.

  • The full details can be provided either with the code, in appendix, or as supplemental material.

Experiment statistical significance

Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments?

Answer: [No]

Justification: Due to computational cost, experiments are reported without multiple runs or error bars. However, improvements are consistent across multiple datasets and architectures.

Guidelines:

  • The answer [N/A] means that the paper does not include experiments.

  • The authors should answer [Yes] if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper.

  • The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions).

  • The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.)

  • The assumptions made should be given (e.g., Normally distributed errors).

  • It should be clear whether the error bar is the standard deviation or the standard error of the mean.

  • It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a 96% CI, if the hypothesis of Normality of errors is not verified.

  • For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g., negative error rates).

  • If error bars are reported in tables or plots, the authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text.

Experiments compute resources

Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments?

Answer: [Yes]

Justification: The paper specifies the hardware setup and training configurations used for experiments (see Appendix).

Guidelines:

  • The answer [N/A] means that the paper does not include experiments.

  • The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage.

  • The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute.

  • The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn’t make it into the paper).

Code of ethics

Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines?

Answer: [Yes]

Justification: The research complies with the NeurIPS Code of Ethics. It involves no human subjects, sensitive data, or high-risk applications.

Guidelines:

  • The answer [N/A] means that the authors have not reviewed the NeurIPS Code of Ethics.

  • If the authors answer [No], they should explain the special circumstances that require a deviation from the Code of Ethics.

  • The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction).

Broader impacts

Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed?

Answer: [N/A]

Justification: This work focuses on foundational methodology for PDE learning and does not have immediate societal deployment implications.

Guidelines:

  • The answer [N/A] means that there is no societal impact of the work performed.

  • If the authors answer [N/A] or [No], they should explain why their work has no societal impact or why the paper does not address societal impact.

  • Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations.

  • The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate Deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster.

  • The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology.

  • If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML).

Safeguards

Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pre-trained language models, image generators, or scraped datasets)?

Answer: [N/A]

Justification: The work does not introduce models or datasets with potential misuse risks.

Guidelines:

  • The answer [N/A] means that the paper poses no such risks.

  • Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters.

  • Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images.

  • We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort.

Licenses for existing assets

Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected?

Answer: [Yes]

Justification: All datasets and prior methods used are properly cited, and their usage follows standard academic practice.

Guidelines:

  • The answer [N/A] means that the paper does not use existing assets.

  • The authors should cite the original paper that produced the code package or dataset.

  • The authors should state which version of the asset is used and, if possible, include a URL.

  • The name of the license (e.g., CC-BY 4.0) should be included for each asset.

  • For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided.

  • If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset.

  • For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided.

  • If this information is not available online, the authors are encouraged to reach out to the asset’s creators.

New assets

Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets?

Answer: [No]

Justification: The paper does not introduce new datasets. Code release is planned but not included in this submission.

Guidelines:

  • The answer [N/A] means that the paper does not release new assets.

  • Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc.

  • The paper should discuss whether and how consent was obtained from people whose asset is used.

  • At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file.

Crowdsourcing and research with human subjects

Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)?

Answer: [N/A]

Justification: The work does not involve human subjects or crowdsourcing.

Guidelines:

  • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects.

  • Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper.

  • According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector.

  1. Institutional review board (IRB) approvals or equivalent for research with human subjects

Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained?

Answer: [N/A]

Justification: The work does not involve human subjects.

  • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects.

  • Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper.

  • We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution.

  • For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review.

  1. Declaration of LLM usage

Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigor, or originality of the research, declaration is not required.

Answer: [N/A]

Justification: LLMs were not used as part of the core methodology.

Guidelines:

  • The answer [N/A] means that the core method development in this research does not involve LLMs as any important, original, or non-standard components.

  • Please refer to our LLM policy in the NeurIPS handbook for what should or should not be described.

References

  1. [1]Matthew W. Abruzzo, Drummond B. Fielding, and Greg L. Bryan. Taming the TuRMoIL: The Temperature Dependence of Turbulence in Cloud–Wind Interactions. The Astrophysical Journal, 966(2):181, May 2024. doi: 10.3847/1538-4357/ad1e51.DOI
  2. [2]Guha Balakrishnan, Amy Zhao, Mert R Sabuncu, John Guttag, and Adrian V Dalca. VoxelMorph: a learning framework for deformable medical image registration. IEEE transactions on medical imaging, 38(8):1788–1800, 2019.
  3. [3]Jungyeon Cho and A. Lazarian. Compressible Magnetohydrodynamic Turbulence: mode coupling, scaling relations, anisotropy, new regime and astrophysical implications. MNRAS, 345:325–339, jan 2003.DOI
  4. [4]Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764–773, 2017.arxiv.org/abs/1703.06211
  5. [5]Bob D De Vos, Floris F Berendsen, Max A Viergever, Marius Staring, and Ivana Išgum. End-to-end unsupervised deformable image registration with a convolutional neural network. In International Workshop on Deep Learning in Medical Image Analysis, pages 204–212. Springer, 2017.
  6. [6]Barath Ezhilan, Michael J Shelley, and David Saintillan. Instabilities and nonlinear dynamics of concentrated active suspensions. Physics of Fluids, 25(7), 2013.DOI
  7. [7]Drummond B. Fielding, Eve C. Ostriker, Greg L. Bryan, and Adam S. Jermyn. Multiphase Gas and the Fractal Nature of Radiative Turbulent Mixing Layers. The Astrophysical Journal Letters, 894 (2):L24, May 2020. doi: 10.3847/2041-8213/ab8d2c.DOI
  8. [8]Tong Gao, Meredith D Betterton, An-Sheng Jhang, and Michael J Shelley. Analytical structure, dynamics, and coarse graining of a kinetic model of an active fluid. Physical Review Fluids, 2(9): 093302, 2017.DOI
  9. [9]Max Gronke and S. Peng Oh. The growth and entrainment of cold gas in a hot wind. Monthly Notices of the Royal Astronomical Society: Letters, 480(1):L111–L115, October 2018. doi: 10.1093/mnrasl/sly131.DOI
  10. [10]Cyrill W Hirt, Anthony A Amsden, and JL Cook. An arbitrary lagrangian-eulerian computing method for all flow speeds. Journal of computational physics, 14(3):227–253, 1974.
  11. [11]Peiyan Hu, Yue Wang, and Zhi-Ming Ma. Better neural PDE solvers through data-free mesh movers. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=hj9ZuNiRm1.arxiv.org/abs/2312.05583
  12. [12]Weizhang Huang and Robert D Russell. Adaptive moving mesh methods, volume 174. Springer Science & Business Media, 2010.DOI
  13. [13]David I. Ketcheson, Kyle T. Mandli, Aron J. Ahmadia, Amal Alghamdi, Manuel Quezada de Luna, Matteo Parsani, Matthew G. Knepley, and Matthew Emmett. PyClaw: Accessible, Extensible, Scalable Tools for Wave Propagation Problems. SIAM Journal on Scientific Computing, 34(4):C210–C231, November 2012.
  14. [14]G.A. Klaasen and W.C. Troy. Stationary wave solutions of a system of reaction-diffusion equations derived from the fitzhugh–nagumo equations. SIAM Journal on Applied Mathematics, 44(1):96–110, 1984. doi: 10.1137/0144008.DOI
  15. [15]Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes. Journal of Machine Learning Research, 24(89):1–97, 2023.
  16. [16]Zongyi Li, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=c8P9NQVtmn0.
  17. [17]Zongyi Li, Daniel Zhengyu Huang, Burigede Liu, and Anima Anandkumar. Fourier neural operator with learned deformations for pdes on general geometries. Journal of Machine Learning Research, 24(388):1–26, 2023.DOI
  18. [18]Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976–11986, 2022.
  19. [19]Huakun Luo, Haixu Wu, Hang Zhou, Lanxiang Xing, Yichen Di, Jianmin Wang, and Mingsheng Long. Transolver++: An accurate neural solver for PDEs on million-scale geometries. In Forty-second International Conference on Machine Learning, 2025. URL https://openreview.net/forum?id=AM7iAh0Krx.arxiv.org/abs/2502.02414
  20. [20]Suryanarayana Maddu, Scott Weady, and Michael J Shelley. Learning fast, accurate, and stable closures of a kinetic theory of an active fluid. Journal of Computational Physics, page 112869, 2024.
  21. [21]GVS Mothish, J Rishi, Shobhit Kumar Shukla, and Deepak Subramani. DNOD: Deformable neural operators for object detection in SAR images. Transactions on Machine Learning Research, 2025. ISSN 2835-8856. URL https://openreview.net/forum?id=tjBqPJDQ72. J2C Certification.
  22. [22]F. Moukalled, L. Mangani, and M. Darwish. The Finite Volume Method in Computational Fluid Dynamics. Springer, 1 edition, 2016. doi: 10.1007/978-3-319-16874-6.DOI
  23. [23]Ruben Ohana, Michael McCabe, Lucas Meyer, Rudy Morel, Fruzsina J Agocs, Miguel Benítez, Marsha Berger, Blakesley Burkhart, Stuart B Dalziel, Drummond B Fielding, et al. The well: a large-scale collection of diverse physics simulations for machine learning. Advances in Neural Information Processing Systems, 37:44989–45037, 2024.arxiv.org/abs/2412.00568
  24. [24]Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International conference on machine learning, pages 5301–5310. PMLR, 2019.arxiv.org/abs/1806.08734
  25. [25]James Rowbottom, Georg Maierhofer, Teo Deveney, Eike Hermann Müller, Alberto Paganini, Katharina Schratz, Pietro Lio, Carola-Bibiane Schönlieb, and Chris Budd. G-adaptivity: optimised graph-based mesh relocation for finite element methods. In Forty-second International Conference on Machine Learning, 2025. URL https://openreview.net/forum?id=pyIXyl4qFx.arxiv.org/abs/2407.04516
  26. [26]Zakhar Shumaylov, Peter Zaika, James Rowbottom, Ferdia Sherry, Melanie Weber, and Carola-Bibiane Schönlieb. Lie algebra canonicalization: Equivariant neural operators under arbitrary lie groups. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=7PLpiVdnUC.arxiv.org/abs/2410.02698
  27. [27]Makoto Takamoto, Timothy Praditia, Raphael Leiteritz, Daniel MacKinlay, Francesco Alesiani, Dirk Pflüger, and Mathias Niepert. Pdebench: An extensive benchmark for scientific machine learning. Advances in neural information processing systems, 35:1596–1611, 2022.arxiv.org/abs/2210.07182
  28. [28]Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, Stéfan J. van der Walt, Matthew Brett, Joshua Wilson, K. Jarrod Millman, Nikolay Mayorov, Andrew R. J. Nelson, Eric Jones, Robert Kern, Eric Larson, C J Carey, Ilhan Polat, Yu Feng, Eric W. Moore, Jake VanderPlas, Denis Laxalde, Josef Perktold, Robert Cimmarman, Ian Henriksen, E. A. Quintero, Charles R. Harris, Anne M. Archibald, Antônio H. Ribeiro, Fabian Pedregosa, Paul van Mulbregt, and SciPy 1.0 Contributors. SciPy 1.0: Fundamental Algorithms for Scientific Computing in Python. Nature Methods, 17:261–272, 2020. doi: 10.1038/s41592-019-0686-2.DOI
  29. [29]Haixu Wu, Huakun Luo, Haowen Wang, Jianmin Wang, and Mingsheng Long. Transolver: A fast transformer solver for pdes on general geometries, 2024. URL https://arxiv.org/abs/2402.02366.
  30. [30]Mingrui Zhang, Chunyang Wang, Stephan Kramer, Joseph G Wallwork, Siyi Li, Jiancheng Liu, Xiang Chen, and Matthew D Piggott. Towards universal mesh movement networks. Advances in Neural Information Processing Systems, 37:14934–14961, 2024.arxiv.org/abs/2407.00382
  31. [31]Zizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable {detr}: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=gZ9hCDWe6ke.arxiv.org/abs/2010.04159

Paper details

vlnc:W_kusv4kty
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents