acceptodds
Under review as a conference paper at ICLR 2027

Geometry-aware latent autoregressive generative model for PDEs in complex domains

Abstract

Solving multiphysics partial differential equations (PDEs) remains a major challenge in scientific computing, especially for highly complex m-scale tortuous geometries critical to energy and chemical engineering. We address this challenge by proposing a Geometry-aware Latent Autoregressive generative Model for PDEs (GeoLAMP), which solves physics within highly irregular and tortuous structures by decoupling flow and transport physics. GeoLAMP introduces a dual-encoder architecture on graph representations to jointly capture global topology and fine-scale geometric features, enabling an effective transition from real-space fields to compact latent representations. In the latent space, we propose a causal self-attention transformer with flow matching to model temporal dynamics, allowing stable and scalable block-wise autoregressive prediction. In addition, we propose a grid–graph data fusion scheme that projects low-resolution grid-based approximate priors onto graph representations, improving prediction of flow in tortuous structures. We establish three multiphysics benchmark datasets in complex geometries, covering reactive flow, heat convection, and elasticity. GeoLAMP consistently achieves the most stable autoregression performance on these datasets. Our results provide a systematic study of geometry-aware learning for PDEs in m-scale complex geometries and offer new insights into block-wise time marching of latent autoregressive PDE modeling via a flow matching framework.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Latent generative model, AI4Physics, PDEs in complex domains

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ckFjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The proposed paper introduces GeoLAMP, a model designed to approximate PDE solution on irregular grids and complex domain. The model is composed of two components: an encoder-decoder based on graph and extracting local and global features, and a flow matching model to performs time stepping. Beyond the model, the paper introduces some datasets with complex geometries. The paper first introduces the model and its different components, and then different experiments are displayed to evaluate the method.

Critical strengths
  • The paper is well-writtten and easy to follow
  • Each component is clearly introduced and illustrated
  • The paper also proposes 3 new datasets for the targeted problem
Critical weaknesses
  • A major concern I have is that main references of the litterature are missing. The idea of making autoregressive steps in a compressed latent space is not new and it has already been proposed in other papers in the past years. Could you comment on the line of works from Kassai et al (ENMA and ECHO [2, 3]) that explored directly the same model structures ? (auto encoder on irregular grid with a time stepping task performs in the compressed latent space using diffusion methods or flow matching)
  • As a result, the main experiment table does not display any recent baselines. While I acknowledge that very recent models can be difficult to spot and reproduce, a two year gap is quite long in the literature (the most recent baseline is Transolver from 2024, why not considering Transolver ++ [1] instead ?). This makes the claim regarding recent baselines comparison not fully demonstrated (line 326).
  • Finally, the novelty of the model is not clear for me or marginal as it relies on a small brick of the encoder.

[1] Transolver++: An Accurate Neural Solver for PDEs on Million-Scale Geometries. Huakun Luo, Haixu Wu, Hang Zhou, Lanxiang Xing, Yichen Di, Jianmin Wang, Mingsheng Long. 2025

[2] ENMA: Tokenwise Autoregression for Generative Neural PDE Operators. Armand Kassaï Koupaï, Lise Le Boudec, Louis Serrano, Patrick Gallinari, 2025.

[3] Efficient Generative Transformer Operators For Million-Point PDEs. Armand Kassaï Koupaï, Lise Le Boudec, Patrick Gallinari. 2025.

Questions for authors
  • Regarding W2, could you compare to more recent model of the literature ?
  • Can you comment on the computational cost of the method ? (memory requirement, training time, inference latency). As the encoder involves computing the distance matrix (in my understanding) in the Global encoder, this operation can be very costly in terms of memory requirements (table 5).
  • Have you trained the model on each dataset separately or all together in a multi physics setting ?
1 comment

Dear Authors, Without any rebuttal provided to answer to my review, I keep my score unchanged.

Reviewer_wuVkOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a method for learning multi-physics PDE emulators where the use of geometry-aware latent autoregression is the novel approach. The approach promises better accuracy particularly when “tortuous” or highly irregular flows exist. By the first paragraph, I am not sure if multi-physics is the primary novelty or if geometry-aware is it, or cross-scale dependence (but it says \mu-m scale or meso-scale). Geometry-aware appears to refer to the combination of global and local information which are learned latent representations.

The paper also introduces three new datasets that are multi-physics (gas and solid; described in the Appendix) and contain the complex geometry they propose to address. Comparisons with other models on these three datasets is provided demonstrating that the proposed model sometimes performs better than others. Ablation studies are conducted and quantified.

Critical strengths

Multi-physics is an under-researched challenge of very high importance in PDE emulation. Furthermore, incorporating the geometry is a promising direction to counteract the standard challenge of predictions “blurring out” over long time horizons. The motivation of the approach is well-founded.

New architectural pieces are introduced by this paper. Notably, the grid-to-graph approach, global vs local encoding (not new, but demonstrating how to implement this successfully), and time block marching (again, not new but important to explore how to overcome significant challenges in long time horizons for PDEs).

Critical weaknesses

The paper’s contributions are unclear from the beginning. By the end of the abstract and first paragraph of the introduction, there are at least 5 different novel solutions promised and so it is unclear what the actual contribution of the paper is (geometry-aware, multi-physics, decoupling flow and transport physics, grid-graph data fusion, establishing 3 benchmark datasets, block-wise time marching, cross-scale computational challenges). Only one of those challenges are mentioned in the clearly-marked “Contributions” paragraph at the end of the Introduction (block-wise time marching, but it is instead called “entire latent block in parallel, with autoregression”). The multi-physics description in Section 3 gives one equation yet claims there are “different physical processes”. More description of the multi-physics and why it is an important challenge that is difficult for existing methods is warranted. Similarly, the results section

Claims are not backed up with evidence. It would be helpful to see comparisons of the model against SOTA on some existing benchmarks. The new model should also perform well on existing benchmarks, as there is plenty of room for improvement on those. Furthermore, it gives a reality-check on how the implementation of those models is performing in known domains. Without these comparisons, we cannot disentangle whether the new model is performing well because of the model, because of the implementation, or because the new datasets are tailored to the new model (and vice-versa). New challenges are introduced, a new model, and new datasets; but the SOTA models compared overlook other models that may already be able to handle these challenges.

Questions for authors

For Table 1, some quantification would help. What are the multiple physics in your datasets? How do you define “complex geometry”? For the new datasets, what are the dimensions of the data and especially how many time steps? A major claim in the paper is handling long time horizons, so it’s important to know what the time horizon is and how it compares to existing datasets. The introduced datasets are the only evaluation, and so understanding the dataset is critical to evaluating the results.

There are foundation models trained on multiple datasets, therefore covering multiple physics (Walrus, POSEIDON, DPOT, MORPH, etc). How do these foundation models perform on your multi-physics benchmarks? Therywere not trained to handle multi-physics in a single dataset, but they have been trained on multiple physics datasets, and so it seems a more fair comparison than against single-physics models like FNO.

In Figure 3, both the top and bottom rows are prediction (according to the caption and they look identical)? The extra column on the right is the only ground truth depiction, yet here the prediction looks vastly different from both ground truth and from the predictions in the preceding columns. Row C has no ground truth? The paper text in Section 5.2 does not discuss the implication of the results shown in Figure 3. Can you shed some light on this figure and what the conclusion from it is?

Replace “casual” typos with “causal”.

Reviewer_kJ2oOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes GeoLAMP, which works in two stages:

  1. A cheap, rough flow solution on a coarse grid is fed to the network, which corrects it into an accurate steady velocity field.

  2. the velocity drives the step-by-step evolution of temperature or chemical concentration.

The paper authors build three new datasets and report the lowest errors against geometry-aware and latent baselines.

Critical strengths
  • New benchmark with whole tortuous domains where there are many holes and internal walls.
Critical weaknesses
  • Splitting flow from transport is exact only because these datasets have one-way coupling (no buoyancy, no geometry change from reaction), so "multiphysics" is somehow overstated.
Questions for authors
  • Does the Table 3 rollout use the predicted velocity from stage 1 or the gt velocity? If gt, how large is the err when both stages run end-to-end?
  • Why does reactive flow with G&L points give 0.0913 and 0.0995 in Table 4, but 0.0789 and 0.0813 for GeoLAMP in Table 3?
  • Is relative L2 computed on raw Kelvin and Pascal, or on normalized fields? How do results change after removing the offset or excluding the near-zero unloading steps?
  • What is the runtime per trajectory vs COMSOL itself and each baseline, including the prior solve?
  • How much do results vary across random noise samples and training seeds?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.