Inverse Chafee–Infante Problem

• The inverse Chafee–Infante problem is the task of reconstructing unknown initial data from near-equilibrium states in a reaction-diffusion system where high-frequency modes decay exponentially.
• Recent approaches employ a physics-informed WGAN-GP architecture with a U-Net generator and forward-simulation penalties to achieve robust, accurate reconstructions.
• Sturm attractor theory provides a rigorous topological framework to classify global attractors and connection graphs, informing the inversion of parabolic PDEs.

The inverse Chafee--Infante problem concerns the recovery of unknown initial data for the Chafee--Infante reaction-diffusion equation, given observations of a near-equilibrium state after significant dissipative evolution. This problem is severely ill-posed owing to the exponential damping of high-frequency spatial modes in the forward dynamics. Recent advances combine physics-informed deep learning architectures—specifically Wasserstein Generative Adversarial Networks with gradient penalty (WGAN-GP)—with explicit residual simulation penalties to achieve stable, accurate reconstructions of initial conditions from evolved data. Parallel notions of Sturm attractor theory in 1D provide a rigorous topological blueprint for classifying global attractors and connection graphs of related parabolic PDEs.

1. Mathematical Formulation of the Forward and Inverse Problems

The Chafee--Infante equation on the square domain $\Omega=[-1,1]^2$ is given by

$$u_t(t,x) - \gamma\,\Delta u(t,x) + \kappa(u(t,x)^3 - u(t,x)) = 0,\quad (t,x)\in (0,T]\times\Omega,$$

subject to Dirichlet boundary conditions $u(t,x)=0$ for $x\in\partial\Omega$, and unknown initial data $u(0,x)=u_0(x)$. The forward map $F$ evolves $u_0$ via 100 explicit forward Euler steps (with $\Delta t=10^{-3}$) to yield $u_{100}(x)=F(u_0)(x)$, a near-equilibrium observation.

The inverse Chafee--Infante problem is to reconstruct $u_0$ given $u_{100}$, i.e.,

$$\text{Given } u_{100},\quad \text{find } u_0 \text{ such that } F(u_0) = u_{100}.$$

Exponentially fast decay of high-frequency eigenmodes (with rates $e^{-\gamma\lambda_k T}$ for Laplacian eigenvalues $\lambda_k$) renders this inversion exponentially sensitive to noise and numerical errors (Shomberg, 12 Jan 2026).

2. Network Architectures for Physics-Informed Inversion

The most effective approach leverages a physics-informed WGAN-GP framework with the following architecture:

• U-Net Generator: Takes $u_{100}\in \mathbb{R}^{128\times128}$ and outputs $\hat u_0\in [-1,1]^{128\times128}$ via an encoder-decoder architecture. The encoder uses five convolutional blocks (feature progression: 64$\to$512), each with 4$\times$4 kernels, stride 2, instance normalization, and LeakyReLU(0.2). The decoder consists of five transpose-convolution blocks, skip connections, and an output $\tanh$ activation.
• PatchGAN Critic: Receives pairs $(u_{100},u)$, where $u$ is either true $u_0$ or $\hat u_0$, and processes them through four convolutional layers (feature progression: 64$\to$512), each with spectral normalization and LeakyReLU(0.2), culminating in a 1$\times$1 convolution and global averaging for the Wasserstein score (Shomberg, 12 Jan 2026).

This architecture is tailored to maintain fidelity to both interfacial structure and global amplitude statistics in the recovered initial state.

3. Physics-Informed Loss Functions

Robust recovery of $u_0$ necessitates a composite loss incorporating both adversarial and physical penalties:

• Wasserstein GAN Loss with Gradient Penalty: Critic maximizes

$$\mathcal{L}_D = \mathbb{E}_{P_g}[D(u_{100},\hat u_0)] - \mathbb{E}_{P_r}[D(u_{100},u_0)] + \lambda_{GP}\,\mathbb{E}_{\hat w}[(\|\nabla_{\hat w}D(\hat w)\|_2 - 1)^2],$$

ensuring smooth Kantorovich potentials. Generator minimizes $-\mathbb{E}_{P_g}[D(u_{100},\hat u_0)]$.

• Lyapunov Energy Matching: Penalizes deviations in the discrete free-energy

$$\mathcal{E}(u) = \int_\Omega \frac{\gamma}{2}|\nabla u(x)|^2 + \frac{\kappa}{4}(u(x)^2 - 1)^2\,dx,$$

enforcing physical interfacial energetics.

• Distributional Statistics Matching: Losses on mean $\mathcal{L}_{\mathrm{mean}}$ and variance $\mathcal{L}_{\mathrm{var}}$ between $\hat u_0$ and $u_0$.
• Forward-Simulation Penalty: Simulates $F^{100}(\hat u_0)$ under the same explicit Euler scheme and penalizes $\|F^{100}(\hat u_0) - u_{100}\|_1$.

Combined, the total generator loss is

$$\mathcal{L}_G = \mathcal{L}_{WGAN} + \lambda_E \mathcal{L}_{\mathrm{energy}} + \lambda_\mu \mathcal{L}_{\mathrm{mean}} + \lambda_\sigma \mathcal{L}_{\mathrm{var}} + \lambda_R \mathcal{L}_{\mathrm{res}} + \lambda_{\mathrm{MAE}}\|\hat u_0 - u_0\|_1,$$

with heuristic weights: $$\lambda_{GP}=10,\,\lambda_E=1.0,\,\lambda_\mu=0.1,\,\lambda_\sigma=0.1,\,\lambda_R=10.0,\,\lambda_{\mathrm{MAE}}=1.0$$ (Shomberg, 12 Jan 2026).

4. Dataset Generation and Numerical Implementation

Training data comprises 50,000 pairs $(u_{100},u_0)$ (test set: 10,000 pairs) generated as follows:

• $u_0(x)\sim \mathrm{Uniform}([-0.02,0.02])$ in the interior, obeying Dirichlet boundary conditions.
• Evolved via explicit forward Euler (100 steps, $\Delta t=10^{-3}$).
• All fields scaled to $[-1,1]$; small Gaussian noise ($\sigma\approx 10^{-3}$) is optionally added to $u_{100}$ for robustness.
• Entire simulation pipeline is implemented using differentiable tensor loops in PyTorch, enabling batched GPU execution. Batch size is 1 due to memory constraints (Shomberg, 12 Jan 2026).

5. Quantitative and Qualitative Results

On the full test set (10,000 pairs), the optimal physics-informed WGAN-GP model achieves

$$\mathrm{MAE}_{\mathrm{test}} \approx 0.23988159,\quad \mathrm{std}(\mathrm{MAE})\approx 0.00266345.$$

Qualitatively, reconstructed $\hat u_0$ closely matches the true interface structure and global amplitude, with minimal artifacts. Even with added noise up to amplitude $10^{-2}$ in $u_{100}$, MAE degrades by less than 5%, indicating strong robustness against ill-posedness and high-frequency damping (Shomberg, 12 Jan 2026).

6. Theoretical Context: Sturm Attractors and Construction Techniques

In 1D, the correspondence between global attractors and connection graphs of semilinear parabolic PDEs is encapsulated by Sturm–meander theory. For equations of the form

$$u_t = u_{xx} + f(x,u,u_x),\quad 0 < x < 1,\quad u_x(0) = u_x(1) = 0,$$

equilibrium profiles are classified by shooting methods, yielding a permutation $\sigma = h_0^{-1} \circ h_1$ identifying boundary height orders. The zero-number $z(u^1-u^2)$ counts sign changes and is strictly decreasing at multiple zeros; Morse indices and adjacency relations are computable via recursive formulas.

A "nose" in permutation encoding $\sigma(j+1)=\sigma(j)\pm 1$ corresponds to a half-circle ("arc") in the meander diagram. The classical Chafee–Infante case features two noses; three-nose meanders are parametrized by coprimality and integer ratios ($p-1$ and $q+1$ coprime, $p=r(q+1)$). Piecewise-Duffing nonlinearities

$$f(x,u,u_x) = \lambda_j u(1-u^2),\quad x\in [x_{j-1},x_j],$$

yield attractors matching prescribed connection graphs when arc lengths $\Delta_j$ match Duffing half-periods $T(\lambda_j)$. Corner smoothings restore $C^2$ regularity (Fiedler et al., 2023).

Global connection graphs $\mathcal{C}$ graded by Morse index reveal lattice unions of Chafee–Infante subgraphs, with an adjoined -1-level vertex $\star$ to maintain connectivity. Removing $\star$ induces a global time-reversibility involution on the directed graph—the heart of the "time-reversible Chafee–Infante lattice" phenomenon.

7. Extensions, Limitations, and Outlook

The physics-informed WGAN-GP approach is extensible to 3D domains, alternative boundary conditions (Neumann, periodic), and more dissipative equations, contingent on computational capacity for forward-simulation penalties. Eyre-type semi-implicit solvers, while unconditionally stable, are unsuitable for gradient-based training due to Newton iteration costs and incompatibility with GPU batching. Methodology generalizes in principle, but further regularization or multi-scale residuals may be required for systems exhibiting stronger dissipation or longer time horizons.

In summary, explicit coupling of adversarial generative modeling with discrete dynamical consistency yields reliable inversion of the Chafee–Infante forward operator. This surpasses classical inverse problem and PINN strategies for reconstructing high-frequency initial data. Time-reversible attractor theory in 1D provides complimentary structural insight into global organization and connectivity of solution spaces (Shomberg, 12 Jan 2026, Fiedler et al., 2023).