# Dirac-Free Weak Mixed PINN

- ID: 2465
- Canonical URL: https://synthcore.org/idea/2465/dirac-free-weak-mixed-pinn
- API JSON: https://synthcore.org/api/idea/2465.json
- API Markdown: https://synthcore.org/api/idea/2465.md
- Verification status: unverified
- Source: [arXiv:2608.17575](https://arxiv.org/abs/2608.17575)
- Category: architecture
- Solves: stability, accuracy, sample-efficiency
- ML areas: loss, architecture, training
- Math tags: pde, functional-analysis, measure-theory, numerical-analysis
- Ratings: usefulness 6/10; difficulty 6/10; novelty 7/10

## Idea description

Build a neural PDE solver that predicts a regularized mixed flux rather than directly fitting a PDE residual containing a Dirac delta. Subtract the explicit radial field generated by the source and train the network with weak constitutive and conservation residuals, so the singularity is represented analytically instead of approximated by a narrow Gaussian.

## Mathematical statement

The paper rewrites the point-source equation as $-\nabla\!\cdot(A\nabla u)=f_{1}+\nabla\!\cdot\mathbf f_{2}$ and defines the modified flux $\boldsymbol\sigma=-A\nabla u-\mathbf f_{2}$, giving $\nabla\!\cdot\boldsymbol\sigma=f_{1}$. In two dimensions, use the explicit source field $\mathbf f_{2}(x;x_{0})=(x-x_{0})/(2\pi\lvert x-x_{0}\rvert^{2})$ for $x\neq x_{0}$; distributionally, $\nabla\!\cdot\mathbf f_{2}=\delta_{x_{0}}$. Here $x_{0}$ is the source location, $A(x)$ is the positive diffusion tensor, $u$ is the potential, and $\boldsymbol\sigma$ is the nonsingular modified flux. Rather than enforce the singular constitutive relation pointwise, use its weak form: for test vector fields $\boldsymbol\tau$, $\int_{\Omega}A^{-1}(\boldsymbol\sigma+\mathbf f_{2})\cdot\boldsymbol\tau\,dx-\int_{\Omega}u\,\nabla\!\cdot\boldsymbol\tau\,dx+\int_{\partial\Omega}g\,\boldsymbol\tau\cdot n\,ds=0$. Conservation is enforced weakly as $-\int_{\Omega}\boldsymbol\sigma\cdot\nabla v\,dx-\int_{\Omega}f_{1}v\,dx+\int_{\partial\Omega}(\boldsymbol\sigma\cdot n)v\,ds=0$. The paper establishes that in the two-dimensional regime the modified flux is naturally in $L^{p}(\Omega)^{2}$ for every $1<p<2$, not generally in $L^{2}$, and that quasi-uniform approximation has flux rate $h^{2/p-1}$.

## Key formulas

- $$-\nabla\!\cdot(A\nabla u)=f_{1}+\delta_{x_{0}}=f_{1}+\nabla\!\cdot\mathbf f_{2},\qquad \mathbf f_{2}(x;x_{0})=\frac{x-x_{0}}{2\pi\lvert x-x_{0}\rvert^{2}},\qquad x\neq x_{0}.$$
- $$\boldsymbol\sigma=-A\nabla u-\mathbf f_{2},\qquad \nabla\!\cdot\boldsymbol\sigma=f_{1}.$$
- $$\mathcal L_{\mathrm{const}}(\theta,\phi)=\sup_{\boldsymbol\tau\in\mathcal T}\frac{\left\lvert\int_{\Omega}\left[A^{-1}(\boldsymbol\sigma_{\theta}+\mathbf f_{2})\cdot\boldsymbol\tau-u_{\phi}\,\nabla\!\cdot\boldsymbol\tau\right]dx+\int_{\partial\Omega}g\,\boldsymbol\tau\cdot n\,ds\right\rvert^{2}}{\|\boldsymbol\tau\|_{L^{p'}(\Omega)}^{2}},\quad p'=\frac{p}{p-1},\quad 1<p<2.$$
- $$\mathcal L_{\mathrm{div}}(\theta)=\sup_{v\in\mathcal V}\frac{\left\lvert-\int_{\Omega}\boldsymbol\sigma_{\theta}\cdot\nabla v\,dx-\int_{\Omega}f_{1}v\,dx+\int_{\partial\Omega}(\boldsymbol\sigma_{\theta}\cdot n)v\,ds\right\rvert^{2}}{\|v\|_{W^{1,p'}(\Omega)}^{2}},\qquad \|\boldsymbol\sigma-\boldsymbol\sigma_{h}\|_{L^{p}}=O\!\left(h^{2/p-1}\right).$$

## Implementation notes

Integrate this at the loss layer of a mesh-free neural PDE solver, not as a replacement for an ordinary Transformer or MLP block. Use two MLPs: $u_{\phi}(x)$ for the scalar potential and $\boldsymbol\sigma_{\theta}(x)$ for the modified flux. Sample interior points $x_i$ from the domain while excluding a small ball $B_\varepsilon(x_0)$; never evaluate $\mathbf f_2$ at $x_0$. Also sample batches of vector and scalar test functions, either fixed compactly supported RBF/Fourier functions or small adversarial test networks $\tau_\psi$ and $v_\omega$. For each batch, compute $\mathbf f_2(x_i)=(x_i-x_0)/(2\pi\|x_i-x_0\|^2)$, then form the weak constitutive residual $R_c(\tau)=\operatorname{mean}[A^{-1}(\sigma_\theta+f_2)\cdot\tau-u_\phi\,\operatorname{div}\tau]$ plus the boundary term. Form the conservation residual $R_d(v)=\operatorname{mean}[-\sigma_\theta\cdot\nabla v-f_1v]$ plus its boundary term. Minimize $\mathcal L=\max_\psi R_c(\tau_\psi)^2/\|\tau_\psi\|_{p'}^2+\max_\omega R_d(v_\omega)^2/\|v_\omega\|_{W^{1,p'}}^2+\lambda_b\operatorname{MSE}(u_\phi-g,\partial\Omega)$. In code, estimate the dual norms by batch averages and update the test networks for two or three ascent steps per primal update. The paper supplies the singular field and the appropriate $1<p<2$ norm; the engineer must estimate only quadrature integrals, boundary traces, and optionally the empirical $L^p$ flux error. Use analytic polar integration or importance sampling for the annulus near $x_0$ because $\|f_2\|$ behaves like $1/r$. First test on $[-1,1]^2$ with $A=I$, $f_1=0$, known Dirichlet data from a logarithmic Green solution, and source points both aligned with and far from collocation points. Compare against a vanilla PINN using a Gaussian delta of widths $\varepsilon\in\{0.1,0.05,0.01\}$. Success means lower seed-to-seed variance, stable training as $\varepsilon$ decreases, and decreasing weak residual plus $L^p$ flux error for $p=1.5$; a useful target is comparable accuracy with at least 2x fewer collocation points and no source-location-dependent failure.

## Disclaimer

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