acceptodds
Under review as a conference paper at ICLR 2027

Bi-FORK: Generative Modeling of High-Dimensional Bifurcating Systems

Abstract

Bifurcations are ubiquitous in physical systems, from structural buckling to fluid and climate dynamics, yet they remain largely unexplored in deep learning. At a symmetry-breaking bifurcation, a single input admits multiple equally valid solutions, violating the one-to-one assumption underlying most learned physical surrogates. We introduce Bi-FORK, a generative framework for learning these one-to-many solution maps in high-dimensional systems. Bi-FORK generates complete trajectories through latent flow matching, preserving space and time coherence, and uses repulsion-guided sampling to recover distinct solution branches in a single amortized pass. We evaluate Bi-FORK on buckling beams, mechanical metamaterials, and Allen–Cahn phase separation, spanning continuous, discrete, and field-valued bifurcations with discretizations up to 260,000 points. Bi-FORK recovers the multimodal solution structure while scaling several orders of magnitude beyond prior approaches, opening generative modeling to high-dimensional bifurcating physical systems.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
flow matching, latent generative model, guidance, symmetry breaking

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_5gJEOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper tackles one-to-many solution maps in symmetry-breaking bifurcations--buckling beams, buckling mechanical metamaterials, and Allen-Cahn phase separation-- where a single condition admits multiple equally valid solutions organized as symmetry orbits.Reported scale reaches 260,000 spatial degrees of freedom, with wall-clock times far below graph-based baselines.

Critical strengths
  1. This is a under-explored and well-posed problem.

  2. MCon-MAE with nearest-orbit matching, per-input JSD, coverage, and rejection rate, plus FFT-descriptor alignment and factorized JSDs for Allen–Cahn, are more appropriate than pointwise CRPS/rank histograms, and the paper explains why.

  3. Full hyperparameter tables, dataset splits with leakage control, and detailed evaluation pipelines.

Critical weaknesses
  1. The closest prior work is neither compared nor adequately differentiated. Hendriks et al. ("Equivariant flow matching for symmetry-breaking bifurcation problems") is the direct predecessor. But the paper never run or compared against.

  2. Particle guidance explicitly optimizes the joint K-sample set for spread. Coverage and per-condition JSD are then computed on exactly that jointly optimized K-sample set. This means the metrics partly measure the objective the sampler was designed to optimize, not the fidelity of the marginal per-sample distribution. I suggest to report marginal metrics computed from independent samples alongside the joint-batch metrics.

3 No analysis of what the latent space preserves. A central, unchecked assumption is that the autoencoder preserves mode identity ("branch identity dominates trajectory variance").

Questions for authors

Please see the weaknesses above.

1 comment

[W1] Model comparison

Thank you for bringing this up. This is a valid point, so we decided to adapt the EGNN-based model by Hendriks et al.[1] for the 2D beams to the 3D beams setting and compare it. The results are the following:

MethodJSDMCon-MAERejection
GeoTDM0.036±0.001\mathit{0.036 \pm 0.001}1.041±0.0011.041 \pm 0.00199.2±0.099.2 \pm 0.0
STFlow0.074±0.0010.074 \pm 0.0010.114±0.0000.114 \pm 0.0000.5±0.10.5 \pm 0.1
Eq. Flow Matching0.026±0.001\mathbf{0.026 \pm 0.001}0.014±0.000\mathbf{0.014 \pm 0.000}0.0±0.00.0 \pm 0.0
Bi-FORK w/o PG0.043±0.0030.043 \pm 0.0030.018±0.001\mathit{0.018 \pm 0.001}0.1±0.10.1 \pm 0.1
Bi-FORK0.041±0.0010.041 \pm 0.0010.018±0.000\mathit{0.018 \pm 0.000}<0.01< 0.01

bold symbols best performing; italic second best performance

The EGNN performs better than Bi-FORK, both in JSD and MCon-MAE. This is mostly likely due to the fact that their model exploits the symmetry, with both equivariance that guarantees an even spread over possible solutions, and the optimal transport approach, which essentially reduces the problem to a 2D problem after all. Just as GeoTDM and STFlow, their model uses graph-based message-passing. By default they all rely on a fully-connected graph, but even when using a different non-fully-connected graph representation, message-passing based models do not scale easily to larger systems (the beams are 3-11 nodes, the metamaterials have up to 36000 nodes). This is also evident from their follow-up work [2] where they apply their framework to the metamaterials dataset. Due to the computational cost, they restrict the model to only use the boundary nodes, and the resulting model is limited in accuracy and speed.

Our intent with Bi-FORK was to go beyond the small test systems in [1] and efficiently scale to larger systems. The difficulty of scaling message-passing approaches is why with Bi-FORK we chose a different approach which effectively uses a latent space. Since we think that the comparison is an overall important addition, we have added the new results to the updated paper and included the motivation for our different approach compared to [1] in Section 3.1. Beam - Results section, page 7.

[W2] JSD metric

We agree that coverage and JSD computed within a jointly guided batch do not alone establish marginal distributional fidelity. We therefore evaluated three settings for the metamaterials, using K=16 per condition: (i) independent unguided sampling, (ii) one jointly guided batch, and (iii) one randomly selected trajectory from each of 16 independently generated guided batches.

SettingCoverage: 2 modes (%)Coverage: 4 modes (%)
(i)36.6±2.136.6 \pm 2.110.7±1.510.7 \pm 1.5
(ii)74.2±3.474.2 \pm 3.449.8±2.949.8 \pm 2.9
(iii)68.2±2.268.2 \pm 2.236.4±1.836.4 \pm 1.8

Setting (iii) also improves branch-frequency JSD over (i), showing that the gain is not solely due to within-batch coordination. Setting (ii) provides an additional set-coverage benefit. These results are drawn using five sampling seeds.

[W3] Branch separability in latent space

Thank you for the question, we agree that reconstruction training alone does not guarantee preservation of branch identity. We therefore additionally evaluated the pretrained autoencoder independently of the generative model on all 445 two- and four-mode metamaterial test conditions. Encode–decode reconstruction retained the original branch in 99.9% of two-mode and 98.6% of four-mode cases. The mismatches only occur when branches are still nearly indistinguishable shortly after bifurcation. These results show that the autoencoder preserves branch identity using the metamaterials benchmark. Thank you for pointing that out, we have added a section to the appendix where we show the results and additional latent space analysis, in Figures 13 and 14, starting from page 27: E.2. Branch separability in latent space. We additionally referenced the section in the introduction: "We evaluate branch preservation through reconstruction in Section E.2."

[1] Hendriks et al., Equivariant flow matching for symmetry-breaking bifurcation problems. arXiv preprint arXiv:2509.03340 (2025b)

[2] Hendriks, F., Equivariant Generative Modeling of Buckling in Mechanical Metamaterials: https://pure.tue.nl/ws/portalfiles/portal/423163048/20261013_Hendriks_hf.pdf (2026)

Reviewer_ZuQ8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses three challenges in high-dimensional bifurcating systems: (1) traditional GNN surrogates fail to scale to systems with tens of thousands of degrees of freedom due to oversmoothing; (2) the "one condition, multiple solutions" problem arising from symmetry breaking requires recovering all branches in a single inference pass. To tackle these issues, the paper proposes three corresponding solutions: first, an autoencoder based on Perceiver/3D-ViT compresses the physical field into a compact latent space, followed by flow matching generation of the entire trajectory in that latent space; second, particle guidance is introduced to apply a repulsive force modulated by a bifurcation frame mask among K concurrent trajectories during sampling. In terms of validation, a three-part evaluation suite is designed, consisting of MCon-MAE (per-mode matched trajectory error), full coverage (branch coverage ratio), and symmetry-orbit-aligned JSD. And on Beam3D the inference speed is approximately 18 times faster than STFlow, demonstrating scalability to high dimensions; on the four-mode Metamaterial experiment, adding PG raises coverage from about 13% to 48%, JSD falls below the random uniform baseline, and visualizations confirm that samples successfully spread across different branches, verifying the effectiveness of parallel deflation.

Critical strengths

By compressing the high-dimensional physical field into a compact latent space via an autoencoder and then performing flow matching generation within that space, the model achieves approximately 18× faster inference than STFlow on Beam3D and can handle the 260,000-point grid of Allen-Cahn, whereas traditional GNN baselines and frame-by-frame generation methods are either infeasible or prohibitively inefficient at this scale.

Critical weaknesses

The paper employs only four simulation realizations along with their sign flips as the set of ground-truth modes. For high-dimensional phase-separating systems, the true branch space is likely considerably larger than this finite collection. Since the computation of JSD depends on pre-defined mode labels, a low JSD value may not necessarily imply that the model has fully recovered all branches, especially if the reference solutions fail to encompass every possible mode. Furthermore, the matching process in MCon-MAE is also constrained by the number of reference solutions, which could lead to an underestimation of the actual error.

Questions for authors

The ground-truth mode set for Allen-Cahn consists of only four reference solutions. We would appreciate it if the authors could clarify how they determined that these four solutions adequately cover all major physical modes. Are there potentially other modes not included, such as patterns with different wavenumbers or phase shifts? If such modes exist, would the reported JSD and coverage values change substantially? Would it be feasible to expand the reference set through additional initial condition sampling or longer simulations?

1 comment

[W1 / Q1] Allen-Cahn Evaluation

MCon-MAE Thank you for the questions regarding the Allen-Cahn evaluation. For the MCon-MAE as mentioned in the paper we do compare the predicted trajectory with its PDE-relaxed trajectory, obtained by a Gauss–Newton projection. It measures the distance to a physically valid solution and does not only compare to the chosen/precomputed modes. To make this more clear we have added a new Figure to the updated main paper, see Section 3.3, page 9, Figure 8.

JSD Evaluation We agree that four modes cannot establish exhaustive branch coverage. The chosen modes are a subset of the problem setting, since generation of a large solution space requires substantial compute. Therefore it is the case that not all patterns within different wavenumbers are covered. But since we generate the 4 modes randomly we assume that our machine learning model can get these different patterns from other inital conditions.

To test how much this matters, we ran a new experiment in which we expanded the reference set from 4 to 32 modes for 20 conditions. We kept the pretrained model fixed and evaluated the same 100 generated samples per condition. Seen in the table below, the mode JSD increases as the mode set grows, while particle guidance consistently achieves a lower mean JSD. Implying that mode JSD depends on the sampled modes and does not establish recovery of all physical branches. Allen–Cahn MCon-MAE is mode-independent since it compares each prediction with its own PDE-relaxed trajectory. Inspired from your questions we have added a section in the Appendix to the new uploaded paper relating this topic, see E.3. Allen-Cahn mode selection, page 29/30.

Modes ∥Mc∥\|\mathcal{M}_c\|Bi-FORK w/o PGBi-FORKLower BoundUpper Bound
40.0980.0800.0090.497
80.1340.1210.0200.574
160.1880.1760.0470.623
320.2570.2440.1070.653

Generating more ground truth data is in principle possible. Nevertheless, we think that at some point they have diminishing returns.

Reviewer_1vzmOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Bi-FORK, a method that employs latent flow matching to generate the complete trajectories of bifurcating systems under specified conditions. It utilizes particle guidance—guided by the predicted bifurcation mask—to improve the coverage of distinct solution branches. The authors validate the proposed approach through applications involving 3D buckling beams, mechanical metamaterials, and Allen–Cahn phase separation, demonstrating its capability to model high-dimensional bifurcating systems.

Critical strengths

The author focuses on bifurcating systems—a significant and common class of complex systems found in the natural sciences. By developing the "Bi-FORK" method, the author overcomes the limitations of existing approaches: the low efficiency of classical numerical methods and the inability of deep learning models to adequately capture multiple potential solutions. The proposed framework—incorporating latent flow matching and particle guidance—is intuitively sound and demonstrates an exceptional capacity for modeling high-dimensional bifurcating systems.

Critical weaknesses

W1. Validation is limited to controlled simulation experiments, lacking tests that better reflect actual bifurcation behavior. Existing results are insufficient to determine the method's robustness and practical applicability under non-ideal conditions.

W2. The authors report only the inference speed of the generative model, omitting the costs associated with data generation and model training. Although the authors argue that these investments can be amortized through repeated inference, the process of generating sufficient training data to cover diverse geometries, parameters, and loading conditions could itself be highly time-consuming; consequently, the overall computational benefit and the conditions required to achieve a "break-even" point remain unclear.

Questions for authors

See the weaknesses part.

2 comments

[W1] Validation

Thank you for bringing up this point. We tried to validate our approach across a broad range of problem settings, going beyond a single controlled experiment. We include continuous beam-buckling directions, metamaterials spanning 17 wallpaper groups with held-out new geometries, and three-dimensional Allen-Cahn dynamics. We chose these settings to test the behavior in controlled settings, to get a proof of concept. Applying the method to real-world examples is the next step, where robustness to imperfections and measurement noise can be tested.

We further discussed this by adding "In this paper, all experiments rely on simulated data, so a natural next step is to test Bi-FORK on experimental measurements, where imperfections and measurement noise influence the solution distribution." to the future work in the Conclusion of the revised paper in blue, on page 10.

[W2] Computation Costs

We agree that the shown inference speed for the Beam3D models alone does not establish an overall computational advantage. The reported timings are important when comparing machine learning models only, since current approaches (GeoTDM, STFlow) lack inference and additionally training efficiency.

Following up on your point we did a full pipeline analysis for the metamaterials and the Allen-Cahn dataset, as the Beam3D dataset can be considered an illustrative toy example. For our computation we follow the breakeven-complexity framework of [1], which is defined as: the up-front cost (data generation plus training) divided by the time the model saves per condition. We report all wall-clock times on B300 and exclude any hyperparameter tuning.

Allen-CahnMetamaterials
Data generation0.75 h346 h*
Training31.7 h68 h
Solver / model per condition3.84 s (4 runs) / 0.64 s (100 samples)112 s* (1 run) / 7.0 s (16 samples)
Speed-up per condition6×16×
Break-even~36,400 conditions~14,200 conditions

* The solver runs on CPU. Simulation times are the mean over all generated samples.

For metamaterials, the one-time cost is about 414 hours, most of it data generation, which can be recovered after about 14,200 conditions. When considering an inverse problem, for example [2] the problem required 30,000 solver runs. For our metamaterials, that would take 39 days with the solver and about 2.5 days with the model. The 36 days saved on this one task are twice the entire one-time cost, and every further task saves the same amount again.

For Allen-Cahn the solver is already very fast, so the model saves little per use and the break-even point is rather far away.

[1] Zhang, Yijing, et al. "Breakeven complexity: A new perspective on neural partial differential equation solvers." arXiv preprint arXiv:2605.15399 (2026).

[2] Li et al., Fourier neural operator for parametric partial differential equations, International Conference on Learning Representations (2021).

Thank you for your reply and revisions. The supplementary results regarding training time has largely addressed my concerns. However, I still have reservations about the statement "every further task saves the same amount again", since it seems that the model has to be retrained from scratch on new tasks. Regardless, I think this is a good job and I'm willing to maintain the current positive evaluation.

Comments on the paper

1 comment

We thank the reviewers for their time, effort and their constructive feedback. We think that the stated questions and outlined weaknesses helped us to improve the quality of our paper. In the revised version of the paper, we highlighted all the changes. This includes:

  1. A comparison with Equivariant Flow Matching on Beam3D, and a clearer explanation of why we chose a latent approach to reach larger systems (Section 3.1, p. 7).
  2. A more thorough Allen–Cahn evaluation. A new figure shows how MCon-MAE uses the PDE-relaxed trajectory (Fig. 8, p. 9), and we expanded the reference set from 4 to 32 modes (Appendix E.3, pp. 29–30).
  3. An analysis of the latent space, showing that the autoencoder keeps the correct solution branch in 99.9% / 98.6% of cases (Appendix E.2, Figs. 13–14, p. 27).
  4. Further evidence for particle guidance, with metrics computed from independent samples rather than a single guided batch (see reviewer 5gJE / [W2]).
  5. A look at practical applicability: a full cost and break-even analysis (see reviewer 1vzm), and validation on experimental data named as a next step (Conclusion, p. 10).

We believe these changes, together with our individual responses, address all of the questions and concerns raised. As the rebuttal period ends soon, please don’t hesitate to let us know if anything is still unclear. We would be more than happy to clarify further.

Thank you for your time and kind regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.