Bi-FORK: A New Generative Modeling Framework for High-Dimensional Bifurcation Systems

In the physical world, bifurcation phenomena are ubiquitous—from structural buckling to fluid and climate dynamics, they describe the process by which a system splits from a single state into multiple possible paths near a critical point. Yet this class of phenomena has long been overlooked in deep learning. Traditional learned physics surrogate models are mostly built on a "one-to-one" mapping assumption: given an input, the model outputs a uniquely determined result. But at symmetry-breaking bifurcation points, one input may correspond to multiple equally valid solutions—a "one-to-many" property that directly challenges the foundations of existing surrogate models.

Recently, a study named Bi-FORK proposed a generative modeling framework for high-dimensional bifurcation systems, attempting to fill this gap. This work not only rethinks how physics surrogate models are constructed from a methodological standpoint, but also validates its effectiveness across multiple high-dimensional scenarios, opening a new direction for applying generative models to physics simulation.

Why Do Bifurcation Systems Break Traditional Surrogate Models?

Bifurcation is a typical manifestation of nonlinear system behavior. When a control parameter crosses a critical value, the system's originally stable single solution may become unstable, and two or more new stable solutions emerge. For example, after the load on a compressed beam reaches a critical point, it may buckle either to the left or to the right—both are physically completely valid.

This symmetry breaking leads to a multimodal solution structure. Traditional supervised learning surrogate models typically train a deterministic mapping, learning the "average" or "dominant" pattern of the training data rather than capturing all possible solution branches. The result: near bifurcation points, the model either produces a compromise solution that does not physically exist, or completely misses certain branches.

The authors of Bi-FORK point out that this absence of one-to-many mapping is the fundamental reason why current deep learning surrogate models perform poorly in bifurcation scenarios.

What is a physics surrogate model? A surrogate model is a common tool in scientific computing and engineering simulation: when first-principles simulations (such as finite element analysis or computational fluid dynamics) are too computationally expensive, researchers train a data-driven approximate model to replace the costly numerical solve. Traditional surrogate models typically adopt deterministic regression or classification architectures (such as multilayer perceptrons, convolutional networks, or neural operators), with the training objective of minimizing the mean squared error between predictions and ground truth—essentially fitting the conditional expectation $\mathbb{E}[y \mid x]$. This design performs well on problems with unique solutions, but when multiple equally valid solutions exist, the conditional expectation "averages" them into an intermediate state that does not physically exist—just as a beam that buckles both left and right would be predicted to remain perfectly straight. Moreover, deterministic models cannot express the uncertainty structure of predictions, and thus cannot distinguish between "a unique solution but model uncertainty" and "the solution itself being multimodal"—two fundamentally different situations.

Bi-FORK's Core Approach: Latent-Space Flow Matching and Repulsion-Guided Sampling

To solve the one-to-many mapping problem, Bi-FORK adopts two key designs.

Latent-space flow matching: The model does not directly generate solutions in physical space, but instead learns a continuous flow in a latent space that transforms a simple prior distribution into the solution distribution. This flow-matching approach preserves the spatial and temporal coherence of solutions, making it especially suitable for generating complete trajectories rather than isolated snapshots.

Repulsion-guided sampling: During inference, Bi-FORK uses a repulsion mechanism to encourage generated samples to spread across different solution branches. In this way, the model can simultaneously recover multiple solution branches in a single amortized inference pass, without needing to train separately for each branch or sample multiple times.

Together, these two designs achieve "one forward pass, multiple solution outputs," in stark contrast to traditional deterministic surrogate models.

Flow matching and continuous normalizing flows. Flow matching is a class of generative modeling methods whose core idea is to learn a time-dependent velocity field $v_\theta(x, t)$ such that samples from a simple prior distribution (typically a Gaussian) can be transformed into the target data distribution after integrating along this velocity field. Unlike diffusion models, which generate samples through progressive denoising, flow matching directly regresses the velocity field, with a training objective typically being the conditional flow matching loss—closer in form to continuous normalizing flows. Its advantages include high sampling efficiency, smooth trajectories, and a natural suitability for handling temporally coherent physical trajectories. In Bi-FORK, latent-space flow matching means the generation process occurs in a low-dimensional latent space rather than the original high-dimensional physical space, which both reduces computational complexity and makes the velocity field easier to learn.

Cross-Scale Validation: From Continuous to Discrete, From Structural to Phase-Field

Bi-FORK's evaluation covers three representative classes of bifurcation systems, corresponding to different types of solution spaces.

• Buckled beam: A classic case of continuous bifurcation, where solution branches correspond to left and right buckling modes.
• Mechanical metamaterials: Discrete bifurcation, where instability modes of unit structures exhibit multistable characteristics.
• Allen-Cahn phase separation: Field-valued bifurcation, involving multiple evolution paths of phase interfaces.

The discretization scale of these cases reaches up to 260,000 points, far exceeding the problem dimensionality that prior related methods could handle. The authors report that Bi-FORK improves scalability by several orders of magnitude while recovering multimodal solution structures. This means that generative modeling can, for the first time, tackle genuinely high-dimensional bifurcating physical systems.

The Allen-Cahn equation and phase-field modeling. The Allen-Cahn equation is a classic model in the phase-field method, used to describe the evolution of phase interfaces in binary alloys or materials. Its form is typically $\partial_t \phi = \epsilon^2 \nabla^2 \phi - W'(\phi)$, where $\phi$ is an order parameter (such as concentration or a phase indicator) and $W(\phi)$ is a double-well potential function. The equation drives the system toward one of two degenerate minima, and thus inherently exhibits bifurcation and multiple solutions: different initial perturbations or boundary conditions lead to different phase interface configurations. Under high-dimensional discretization, the Allen-Cahn system can produce a large number of spatial points (such as the 260,000 points in this paper), and its solution space is highly multimodal. Bi-FORK chose this case precisely to validate the scalability of generative surrogate models on field-valued bifurcation problems.

Significance for Physics Simulation and Generative AI

The value of Bi-FORK lies not just in a specific model, but in redefining the role of deep learning surrogate models in scenarios with non-unique solutions.

From an engineering perspective, bifurcation analysis is crucial in fields such as structural stability, materials design, and climate simulation. A surrogate model capable of generating all possible solution branches at once can significantly accelerate uncertainty quantification and robustness assessment.

From a methodological perspective, Bi-FORK pushes generative modeling further from domains like images and text into scientific computing. It shows that techniques such as flow matching and guided sampling are equally applicable to high-dimensional dynamical systems with strict physical constraints.

Of course, this framework is still at the research stage. How to ensure that generated solutions strictly satisfy physical conservation laws, how to quantify the completeness of branch coverage, and how to extend to more complex spatiotemporal chaotic systems are all questions worth exploring in future work. Regardless, Bi-FORK provides a clear and scalable starting point for generative modeling of high-dimensional bifurcation systems.

One-line Summary

Bi-FORK uses latent-space flow matching and repulsion-guided sampling to let a surrogate model output multiple solution branches of a high-dimensional bifurcation system in a single forward pass.

Paragraph Summary

Traditional learned physics surrogate models are built on a "one-to-one" mapping assumption and fail near symmetry-breaking bifurcation points, because one input may correspond to multiple equally valid solutions. Bi-FORK proposes a generative modeling framework that learns the solution distribution via latent-space flow matching and uses repulsion-guided sampling to encourage samples to spread across different branches, enabling recovery of multiple solutions in a single amortized inference pass. The method is validated on three classes of bifurcation systems—buckled beams, mechanical metamaterials, and Allen-Cahn phase separation—with discretization scales reaching up to 260,000 points, improving scalability by several orders of magnitude over prior methods. This provides a new paradigm for generative modeling of high-dimensional bifurcating physical systems and holds promise for accelerating uncertainty quantification in structural stability, materials design, and climate simulation.