acceptodds
Under review as a conference paper at ICLR 2027

Aster: Learning Tensor-Network Structure for PDE Solving with a Differentiable Quantum–Classical Model

Abstract

Tensor networks offer compact representations of high-dimensional fields and are increasingly used for solving partial differential equations (PDEs) at resolutions where conventional discretisations become computationally prohibitive. However, most tensor-network PDE methods rely on fixed structures chosen in advance, while existing structure-search methods primarily optimise compression of individual fields rather than PDE dynamics. Although quantum circuits provide flexible trainable samplers, their use for tensor-network structure discovery is further limited by the lack of efficient mechanisms for generating only valid network structures. We introduce Aster, a hybrid quantum–classical model that jointly learns tensor-network structures and PDE dynamics. A trainable quantum sampler generates valid network topologies without rejection and integrates them into a differentiable tensor-network solver. Aster further incorporates rank adaptation into the discovered network structure. Across three evaluation settings, Aster matches the accuracy of spectral Runge–Kutta on six of seven benchmark PDEs while using 2.9–10 fewer stored scalars on five. Aster-Solver achieves the lowest error at every evaluated resolution on eight of twelve two-dimensional PDEs against state-of-the-art neural PDE solvers, while using substantially fewer parameters. As a structure-search method, Aster also achieves lower storage at matched accuracy than existing approaches at the three highest resolutions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Tensor Networks, Architecture Search, PDE Solvers

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_gZJwOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, the authors proposed ASTER, a differentiable quantum–classical framework for learning tensor-network structures for PDE solving. A trainable quantum sampler generates valid network topologies without rejection, and a differentiable tensor-network solver jointly learns the representation and PDE dynamics. An extended variant further incorporates rank adaptation to adjust representation capacity. Experimental results on various tasks, including structure search, demonstrate competitive accuracy with reduced storage and parameter requirements.

Critical strengths
  1. The proposed method has good motivation. Jointly adapting tensor-network topology and PDE dynamics addresses a limitation of fixed-topology solvers and snapshot-based compression methods.
  2. The bit-to-tree decoder has a clear validity guarantee under the stated noiseless, number-preserving circuit assumptions. Combining topology search with rank adaptation provides a flexible representation.
Critical weaknesses
  1. The quantum-sampler comparison includes a classical uniform sampler, but this does not isolate the value of the quantum parameterization from the benefit of learning a non-uniform topology distribution.
  2. Reduced storage and parameter counts do not directly imply lower total computational cost. Sampler training, repeated PDE evaluations, and quantum-gradient estimation can add substantial overhead.
Questions for authors
  1. How does the quantum sampler compare with a trainable classical autoregressive or categorical sampler using the same valid decoder, solver, and search budget?
  2. What are the end-to-end training and inference times, including topology search and gradient estimation, at matched solution accuracy?
  3. How does the solver behave over longer integration horizons and unseen initial conditions? Please examine conservation errors and stability when topology or rank changes.
Reviewer_aDNxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ASTER, a quantum–classical framework that uses a trainable quantum circuit to search for tensor network and further combines HSVD, tangent-projected time integration, and a learned closure for PDE solving. The author further proposes ASTER+ by enabling adaptive rank selection and ASTER-Solver extending the framework to data-driven PDE prediction. Experiments show that the proposed methods reduce storage while maintaining accuracy compared to classical numerical solvers. ASTER-Solver achieves lower errors than neural operator baselines on several time extrapolation data-driven benchmarks.

Critical strengths

It is interesting to use quantum circuit to approach tensor network structure search. The sampling and decoding guarantees valid tree structures, providing a trainable framework for exploring different tensor network structures.

Critical weaknesses
  1. Limited practical significance. The main benefit of the proposed method is the reduced storage, but the numerical experiments are limited to small problems (2D (32232^2) and 3D (16316^3)), where memory is unlikely to be bottleneck. Moreover, all seven PDEs in these experiments have closed-form solutions and do not necessarily need numerical solutions. It is also unclear how much computational overhead this method introduces during training, comparing with numerical solvers where no training is needed. In terms of accuracy, the proposed method also does not beat the existing numerical methods.
  2. Unclear quantum advantages. The paper does not establish a clear advantage of quantum sampling over classical sampling. While the authors compare their approach with classical uniform sampling, a comparison with trainable classical structure samplers is still needed to prove the advantage of the quantum sampler. Furthermore, all quantum experiments in this paper rely on noiseless MPS simulation, and the impact of quantum hardware noise is not considered.
Questions for authors
  1. Could the authors provide the total training time and computational cost of the proposed method?
  2. FNO and DeepONet perform surprisingly poorly on several relatively simple PDEs. Could the authors clarify whether these baselines were sufficiently trained and properly tuned?
Reviewer_Y8bQOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ASTER, a hybrid quantum-classical framework for learning tensor-network structures during PDE solving. The motivation is that most existing tensor-network PDE methods use fixed topologies, while structure-search approaches primarily focus on compressing individual fields rather than representing evolving PDE solutions.

The method uses a parameterized quantum circuit to sample tensor-network topologies, mode orderings, and bond dimensions. A bit-to-tree decoder guarantees valid structures without rejection. These structures are then used in a differentiable tensor-network solver that projects PDE dynamics onto the corresponding low-rank manifold, with a learned correction term. ASTER+ further incorporates adaptive tensor ranks, while ASTER-SOLVER is designed for data-driven PDE prediction.

The experiments cover numerical integration, learned PDE solving, and tensor-network structure search. The authors report competitive accuracy with substantially reduced storage on several PDE benchmarks, and improvements over existing tensor-network structure-search methods at higher resolutions.

Critical strengths
  1. Learning tensor-network structures for evolving PDE solutions rather than fixing them beforehand is an interesting direction. The combination of topology search, rank adaptation, and differentiable PDE integration is reasonably motivated and could be useful for high-dimensional problems where dense representations become expensive.
  2. The proposed bit-to-tree decoder guarantees valid tree structures without rejection. The construction is mathematically well motivated, and the theoretical result provides a useful guarantee for structure generation.
  3. The experiments cover several aspects of the method, including numerical integration, learned PDE solving, and tensor-network structure search. The reported storage reductions at comparable accuracy are promising, particularly at higher resolutions. The authors also acknowledge some cases where the method does not perform well.
Critical weaknesses
  1. I am not fully convinced that the quantum sampler is necessary or particularly advantageous for this problem. The main benefit of the proposed decoding scheme is that it maps samples to valid tensor-network structures, but this construction is not inherently quantum and could also be used with classical samplers. The paper does not establish what specific capability of the quantum circuit makes it preferable to a classical learned distribution over the same structures.
  2. Related to the previous point, the classical sampler comparison in Table 9 is rather limited. The quantum sampler is trainable, while the classical alternative uses uniform sampling. This is not a particularly strong comparison, since the improvement could simply come from learning the sampling distribution rather than from the quantum parameterization. A trainable classical sampler, such as an autoregressive or categorical model using the same decoder and optimization objective, would be a more informative baseline. This is especially important since the authors acknowledge that the quantum circuits used in the experiments are efficiently classically simulable.
  3. The framework combines several components, including topology search, rank adaptation, projected integration, and a learned correction term. While the overall results are promising, it is difficult to determine how much each component contributes. In particular, a fixed-topology tensor network with the same rank adaptation and learned closure would be a useful comparison. Without this, the necessity of adaptive topology search itself is not sufficiently established.
  4. The paper emphasizes reduced storage, but the overall computational efficiency is less clear. Structure sampling, tensor-network truncation, projected integration, and joint optimization all introduce additional computational costs. The number of stored scalars alone does not capture these costs. More direct comparisons of runtime and training overhead would help establish whether the method offers a practically useful tradeoff. The different structure-evaluation budgets used in the structure-search experiments also make the comparison less straightforward.
  5. Some of the experimental evidence is also limited by the number of runs. In particular, the neural PDE solver comparisons appear to use only one seed. Given the stochastic structure sampling and training procedure, it is difficult to assess the stability of the reported improvements. This matters especially where the accuracy differences are relatively small.
Questions for authors
  1. What specific advantage does the quantum sampler provide over a trainable classical distribution using the same bit-to-tree decoder? Are there any results with a learned classical sampler, or evidence that the quantum parameterization provides a better search distribution under comparable budgets?
  2. Could the authors clarify how much of the improvement comes from topology adaptation itself? In particular, is there a comparison with a fixed topology while keeping rank adaptation, the learned closure, and the remaining training procedure unchanged?
  3. How does ASTER compare with the classical structure-search methods in terms of total runtime and number of structure evaluations? Could the authors clarify how the different evaluation budgets affect the reported comparison?
  4. Are results from multiple random seeds available for the ASTER-SOLVER experiments? It would be helpful to understand the variability introduced by the stochastic sampler and joint optimization.
1 comment

Once again, we sincerely thank you for your valuable comments. We hope that our responses have adequately addressed all of your concerns. We would be happy to address any further questions or provide additional clarification regarding the novelty and contributions of our work.

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

The paper proposes ASTER, a framework for learning tensor-network structures for PDE solving. A parameterized quantum circuit samples the tree topology, mode ordering, and bond dimensions, while a classical solver combines hierarchical SVD truncation, projected time integration, and a learned correction to the dynamics. A cycle-lemma decoder ensures valid tree shapes without rejection, and ASTER+ additionally adapts ranks during integration. The experiments evaluate numerical integration, learned PDE prediction, and compression, reporting reduced storage at comparable accuracy and improved prediction relative to several neural baselines.

Critical strengths

1.The rejection-free tree decoder is a clean contribution. The cycle-lemma construction maps the sampled fixed-Hamming-weight strings to valid binary trees, and the argument is straightforward to check. This addresses a concrete difficulty in topology search. Searching over the tree and mode ordering together is also a sensible design, since both influence which correlations the tensor network can represent efficiently.

2.The rank-adaptive integration results are promising. The matched-accuracy comparison in Table 7 is particularly useful: on several equations, ASTER+ retains accuracy comparable to spectral RK4 with substantially fewer stored scalars. This supports the practical value of adapting the representation during evolution. The paper also reports cases where compression does not help and distinguishes core storage from runtime and peak memory.

Critical weaknesses
  1. The rank decoder does not guarantee compatibility with the fixed-rank formulation. My main technical concern is that Algorithm 7 enforces dimension bounds on each edge independently, whereas Section 3.1 defines MS\mathcal{M}_S using exact ranks. Consider a balanced four-leaf tree. The two edges connecting the root to its children correspond to complementary matricizations of the same tensor, so their ranks must agree. Algorithm 7 nevertheless allows ranks 1 and 2 on these edges. No tensor has that pair of exact ranks.

    This matters for the interpretation of the truncation and tangent projector in Eq. (2). If χ\chi is intended to specify capacity bounds, the paper needs to explain how the realized ranks are determined and how the projector handles bounds that are not attained. The example directly concerns the sampled-rank ASTER construction. ASTER+ recomputes ranks, so its results are not necessarily affected, but the relevant implementation paths need to be checked.

  2. The reported experiments are not clearly connected to the claimed joint training procedure. Section 3.3 describes a single training loop and backward pass, and Appendix A.3 includes a joint stage using the rollout objective. In the supplied code, however, dynamics/solver.py uses a NumPy-based rollout, while FieldNet in baselines/neural.py is fitted separately on (u,du/dt)(u,\mathrm{d}u/\mathrm{d}t) pairs and evaluated under torch.no_grad(). These may be evaluation-only paths, but I could not locate the training path that updates the closure through Eq. (4). Given the additional use of an initial-state compression proxy for structure search, it remains unclear which results demonstrate joint learning and what that stage contributes.

    The neural-baseline protocol also needs clarification. Table 1 specifies 100 optimizer steps with batch size 8 for the operator baselines and 250 steps with batch size 128 for the closure/surrogate, whereas Appendix B.2 states that all methods are trained for 150 epochs. DeepONet's relative errors are also close to one on many tasks in Table 6. Different methods need not receive identical update budgets, but these comparisons require a consistent protocol and evidence that the baselines were adequately optimized.

Questions for authors
  1. How is rank compatibility enforced before constructing the projector? Please address the four-leaf example above, clarify whether χ\chi denotes exact ranks or upper bounds, and explain whether a correction changes any reported results.

  2. Which implementation produced the ASTER-SOLVER results? Please identify the path that trains the closure using Eq. (4), explain how the sampler is updated, and specify which experiments use the joint stage in Appendix A.3.

  3. What training budgets and supervision were actually used for the neural comparisons? Please reconcile Table 1 with Appendix B.2 and provide the available training/validation evidence for adequate optimization, particularly for DeepONet.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.