# Smooth-Plus-Boundary-Lifting Network

- ID: 361
- Canonical URL: https://synthcore.org/idea/361/smooth-plus-boundary-lifting-network
- API JSON: https://synthcore.org/api/idea/361.json
- API Markdown: https://synthcore.org/api/idea/361.md
- Verification status: unverified
- Source: [arXiv:2607.02312](https://arxiv.org/abs/2607.02312)
- Category: architecture
- Solves: accuracy, stability, sample-efficiency
- ML areas: mlp, architecture, loss, training
- Math tags: pde, functional-analysis, approximation-theory
- Ratings: usefulness 5/10; difficulty 5/10; novelty 8/10

## Idea description

Represent the prediction as a sum of a smooth interior branch and a fractional boundary branch: u_theta(x)=u_int_theta(x)+d(x)^a u_bd_theta(x). This mirrors the paper's direct-sum solution structure and allocates separate network capacity to the globally regular component and the boundary layer.

## Mathematical statement

For forcing data of regularity s in [0,tau-2a), the paper describes the solution space as a direct sum of a full-regularity component H_q^(2a+s)(closure Omega) and a component d^a times a Poisson lifting of boundary values. Here 2a is the order of P, d(x)=dist(x,partial Omega), and tau describes boundary regularity. The proposed network uses u_int_theta to model the full-regularity component and d^a u_bd_theta to model the boundary-lifting component. The PDE residual trains their sum, while branch regularizers encourage u_int_theta to remain smooth and u_bd_theta to act primarily near the boundary.

## Key formulas

- $$\mathcal S_f=\dot H_q^{2a+s}(\overline\Omega)\oplus d^a\mathcal L_s,$$
- $$u_\theta(x)=u^{\mathrm{int}}_\theta(x)+d(x)^a u^{\mathrm{bd}}_\theta(x),$$
- $$\mathcal L_\theta=\operatorname{mean}_{x_i\in X}|P[u_\theta](x_i)-f(x_i)|^2+\lambda_{\mathrm{sep}}\left(\|\nabla u^{\mathrm{int}}_\theta\|_2^2+\|d^{1-a}\nabla u^{\mathrm{bd}}_\theta\|_2^2\right).$$

## Implementation notes

Use two coordinate networks with shared low-level features: N_int_theta(x) for the smooth branch and N_bd_theta(x) for the boundary quotient. Define u_theta(x)=N_int_theta(x)+d_delta(x)^a N_bd_theta(x), using a smoothed distance d_delta when needed, and set u_theta=0 outside the domain. At every training step, sample ordinary interior points X and oversample a boundary strip X_b={x:d(x)<rho}. Evaluate the combined solution, compute r=P[u_theta]-f, and minimize the PDE residual plus a branch regularizer. The exact paper ingredient is the direct-sum structure H_q^(2a+s) plus d^a times a lifting. In code, approximate smoothness with automatic-differentiation gradients evaluated on X, and approximate boundary-branch localization with m_b(x)=exp(-d(x)/rho). Use L_sep=mean(||nabla N_int||^2)+lambda_bd mean(m_b||d_delta^(1-a)nabla N_bd||^2)+lambda_far mean((1-m_b)N_bd^2). Initialize the interior branch with low-frequency Fourier features and the boundary branch with ordinary positional features. Test on a unit disk or square using manufactured solutions containing both a smooth term and d^a times a smooth term. Compare with a single unconstrained PINN and the single factored architecture, keeping parameter count and collocation budget fixed. Track total L2 error, boundary-strip error, residual convergence, and branch magnitudes away from the boundary. Success means lower boundary-layer error at equal parameter count without worsening interior accuracy.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.