Feature-Based Continuation of Pattern Transitions in a One-Dimensional Brusselator

Abstract: Different long-time patterns can prevail in different regions of a reaction--diffusion system's parameter space. We study the transition curves between such regimes for a one-dimensional Brusselator in the two-parameter plane $(σ,b)$, focusing on wave/stripe-like, spiral/source-defect-like, and target-like states. We develop a feature-based continuation framework built on time-dependent PDE simulations. Scalar observables extracted from late-time solution data distinguish the regimes and define threshold level sets where the feature crossings are regular. A secant predictor and a local one-dimensional sweep corrector are used to trace these level sets. For the spiral transition, we introduce a branch-adapted spacetime symmetry-defect feature that separates asymmetric source-like patterns from more symmetric wave patterns. For the target transition, we use the minimum of a core spatial-variance score and a tail temporal-variance score to detect the characteristic structure of half-target states. The method recovers robust side portions of both transition curves. In lower parameter regions, where mixed and irregular patterns make a single scalar feature less specific, we instead report transition estimates obtained from vertical parameter sweeps and direct inspection of spacetime plots. These results show how simulation-based continuation and direct pattern classification can be combined to map regime boundaries while preserving the different levels of numerical evidence.

• The paper develops a feature-based predictor–corrector framework that traces regime boundaries as threshold level sets of observables extracted from time-dependent Brusselator simulations.
• It introduces branch-aware crossing rules, hysteresis, cold-start detection sweeps, and specialized symmetry and target detectors to improve continuation reliability near curved or narrow transitions.
• The results recover robust upper and side branches, but mixed localized patterns, detector dependence, threshold choices, and corrector geometry limit confidence in lower-boundary segments.

This paper develops a feature-based continuation framework for mapping regime boundaries in a one-dimensional Brusselator reaction–diffusion model, formulated in the two-parameter plane $(\sigma,b)$ where $b$ is the kinetic parameter and $\sigma$ sets the diffusion ratio $d_1=\sigma d_2$ (with $a=2.5$, $d_2=9.73$, homogeneous Neumann conditions). Rather than continuing exact coherent structures via boundary-value formulations, the work treats transition curves between late-time pattern regimes as threshold level sets of scalar observables extracted from time-dependent PDE simulations, following the data-driven continuation philosophy of Zhao, Maffa, and Sandstede (Zhao et al., 19 Feb 2025).

Problem setting and motivation

The model is the classical two-component Brusselator PDE solved by method of lines on a uniform grid, integrated with ode45 in MATLAB, with the second component $v(t,x)$ stored as a spacetime matrix used for all feature extraction. Three qualitative late-time regimes motivate the study: wave/stripe-like states (regular, symmetric wave trains), spiral/source-defect-like states (localized cores emitting waves asymmetrically, interpreted as one-dimensional analogues of the "spirals" observed experimentally by Perraud et al.), and target-like states (a structured core with an oscillatory tail, related to pacemaker theory). The author is careful to note that these are late-time numerical regimes rather than analytically characterized solution classes, and that one-dimensional "spirals" are source-like analogues, not literal two-dimensional rotating spirals.

The central methodological claim is that direct parameter scans are too coarse near narrow or curved transition regions, while classical continuation requires an explicit bifurcation equation that is unavailable here. The compromise is to define each transition as $$\Gamma = \{(\sigma,b): r(\sigma,b)=r_{\mathrm{thr}}\}$$ for a simulation-derived observable $r$, and to trace this level set numerically.

Continuation algorithm

The continuation uses a secant predictor along the local tangent direction followed by a local one-dimensional sweep corrector: fixing $b=b^{\mathrm{pred}}$, the algorithm samples $b$0 locally, evaluates the feature from fresh PDE simulations, and selects a threshold crossing. Several refinements prove essential:

• Branch-aware crossing direction: crossings must satisfy a sign condition on $b$1, since different branches cross the threshold in opposite directions.
• Hysteresis: among multiple admissible crossings, the one nearest the previously accepted point is preferred, regularizing branch continuity.
• Warm-start versus cold-start: warm-starting is used along the accepted branch but explicitly avoided within detection sweeps, since warm-started sweeps make the sampled feature path-dependent rather than a well-defined function of $b$2.

A temporal-coherence diagnostic, based on mean adjacent-frame differences of the late-time spacetime field, flags irregular states for direct inspection. The method is positioned as inspired by pseudo-arclength continuation but not equivalent to it: the corrector is a simulation-based threshold search, not a Newton solve of an augmented system.

The spiral feature

An earlier final-profile symmetry metric proved too weak because two states can share similar terminal profiles while differing in spacetime organization. The adopted detector instead measures a normalized Frobenius-norm mismatch between the late-time spacetime field (last 25% of the observation window) and a reflected copy. Notably, the reflection operator is branch-dependent: space reflection ($b$3) works cleanly on right branches, whereas it produces large spurious spikes on left branches, where time reflection ($b$4) is robust. The author emphasizes this is an empirically validated choice, not a universal property of the PDE.

The target feature

The target detector is conjunctive: $b$5, where $b$6 is the time-averaged spatial variance over an interior core window $b$7 and $b$8 is the space-averaged temporal variance over a tail window $b$9, both computed on the last 40% of the observation window. The minimum enforces a logical "and" — both a structured core and an oscillatory tail must be present — preventing one large component from masking the absence of the other. The continuation threshold is fixed empirically at $\sigma$0, chosen because target-like states lie comfortably above it across representative sweeps at $\sigma$1, $\sigma$2, and $\sigma$3. Left and right target branches are tracked by upward and downward crossings respectively.

Numerical results

For both transitions, the framework recovers robust upper and side branches of the regime boundary in the $\sigma$4-plane. In lower parameter regions, however, two difficulties emerge. First, geometric: the lower boundary behaves like $\sigma$5 rather than $\sigma$6, so the horizontal corrector is misaligned. Second, and more fundamentally, mixed or irregular localized-source patterns produce threshold crossings that do not correspond to the intended coherent transition. A vertical-corrector test produced a smooth symmetry-threshold curve that continued through the mixed region without joining the visually identified boundary — a concrete demonstration that corrector geometry alone does not guarantee detector validity. Consequently, lower middle-left and middle-right segments are reported only as vertical-sweep estimates bracketed by visual classification of spacetime plots, with connecting lines labeled explicitly as interpolation guides rather than continuation output.

Validation is external to the continuation loop: continued side branches were checked against independently computed horizontal sweeps, and for the target right-up branch the sweep-derived transition points align closely with the continuation points, supporting the interpretation of the branch as a genuine feature-defined boundary rather than an artifact of predictor–corrector logic. This validation is inherently local, however, and does not extend to the lower boundary.

Limitations

The paper is candid about several limitations. Detector dependence is primary: continuation quality is bounded by feature specificity, as shown by the failure of the final-profile metric and the need for branch-adapted reflections. Threshold dependence follows: the level-set location depends on empirically chosen thresholds not determined by theory. The one-dimensional corrector is sensitive to local curve orientation, and the overall pipeline is not autonomous — hysteresis windows, crossing rules, coherence checks, and visual inspection all inject judgment. Finally, some ambiguity originates in the dynamics itself: the PDE supports grey-zone patterns that no scalar detector can classify without stronger modeling assumptions.

Conclusion

The paper demonstrates that predictor–corrector continuation can be adapted to feature-defined transition problems when no explicit boundary-value formulation exists, provided the different tiers of numerical evidence — continued level sets versus visually bracketed sweep estimates — are kept distinct. Its main open questions concern whether adaptive or full two-dimensional correctors can resolve the lower boundaries, whether refined features can sharpen specificity in mixed regions, and whether the computed curves correlate with analytically identifiable Turing–Hopf instability thresholds or codimension-two interactions in the Brusselator.