# Artificial-Compressibility Divergence Feedback

- ID: 2968
- Canonical URL: https://synthcore.org/idea/2968/artificial-compressibility-divergence-feedback
- API JSON: https://synthcore.org/api/idea/2968.json
- API Markdown: https://synthcore.org/api/idea/2968.md
- Verification status: unverified
- Source: [arXiv:2609.00216](https://arxiv.org/abs/2609.00216)
- Category: dynamics
- Solves: stability, accuracy, training-dynamics
- ML areas: world-model, graph-nn, training-dynamics, loss
- Math tags: pde, dynamical-systems, control-theory, differential-geometry
- Ratings: usefulness 6/10; difficulty 5/10; novelty 6/10

## Idea description

Add a pressure-like recurrent state to a neural surface-flow decoder and update it from the predicted local divergence, creating a learned or fixed feedback loop that drives vector outputs toward local incompressibility. Unlike a static divergence penalty, the state can accumulate constraint violations and produce corrective tangent gradients at each refinement step.

## Mathematical statement

The paper uses an artificial-compressibility/EDAC pressure equation containing the feedback term \(-\mathrm{Ma}^{-2}\nabla\cdot\boldsymbol v\), where \(\boldsymbol v\) is surface velocity, \(\nabla\cdot\boldsymbol v\) is intrinsic surface divergence, and \(\mathrm{Ma}>0\) is an artificial Mach number controlling stiffness. The paper interprets this term as a local feedback controller: negative divergence indicates accumulation and changes pressure so that the pressure gradient pushes the flow outward. The incompressible target is \(\nabla\cdot\boldsymbol v=0\). The neural adaptation introduces a discrete pressure state \(p^k\), computes a tangent pressure gradient \(g^k=\nabla p^k\), and corrects the network velocity by subtracting \(g^k\). The explicit loop transfers the paper's local constraint-feedback mechanism into iterative refinement of a neural prediction.

## Key formulas

- $$\nabla\cdot\boldsymbol v=0,$$
- $$\text{EDAC feedback term}=-\frac{1}{\mathrm{Ma}^{2}}\nabla\cdot\boldsymbol v,$$
- $$p^{k+1}=p^k-\frac{\Delta\tau}{\mathrm{Ma}^{2}}\,D(v^k),$$
- $$v^{k+1}=v^k-\Delta\tau\,G(p^{k+1}),\qquad D(v)\approx\nabla\cdot v,\quad G(p)\approx\nabla p.$$

## Implementation notes

Use this in a neural operator or graph network that predicts a tangent velocity field on a surface point cloud. Let the backbone produce an initial velocity \(v^0_i\) and initialize a scalar pressure state \(p^0_i=0\). Build fixed meshfree differential operators from local neighborhoods: estimate normals and projectors \(P_i\), fit local tangent coordinates by PCA, and compute divergence \(D(v)_i\) and scalar gradient \(G(p)_i\) by weighted least-squares polynomial reconstruction. For refinement steps \(k=0,\ldots,K-1\), execute: first compute \(d^k=D(v^k)\); then update \(p^{k+1}=p^k-(\Delta\tau/\mathrm{Ma}^2)d^k\); finally update \(v^{k+1}=P[v^k-\Delta\tau G(p^{k+1})]\). Optionally apply a small residual network \(v^{k+1}\leftarrow P[v^{k+1}+R_\theta(v^{k+1},p^{k+1})]\). Train against the target velocity with \(L=\|v^K-v^*\|^2+\lambda\|D(v^K)\|^2\), but test whether the recurrent feedback itself, rather than the penalty, provides the benefit. Estimate the stable range by sweeping \(\alpha=\Delta\tau/\mathrm{Ma}^2\); do not assume the physical PDE timestep is stable for a learned decoder. The first experiment should use a small graph network on analytically generated divergence-free tangent fields on sphere and torus point clouds, with 1, 2, 4, and 8 refinement steps and matched FLOPs against a feed-forward baseline. Pre-register that final divergence decreases approximately monotonically with \(K\) for an intermediate \(\alpha\); too-large \(\alpha\) produces oscillation or divergence growth; and, at fixed final divergence, the feedback model requires at least 25% less explicit divergence-penalty weight or fewer refinement steps than the static-penalty baseline. Measure the dominant temporal oscillation frequency of a tracked point or mode; the EDAC mechanism predicts that increasing \(\mathrm{Ma}\) lowers the feedback-induced frequency, approximately following the inverse-Mach trend. Ablate only the pressure state and gradient correction while retaining the same backbone and divergence loss. Falsify the idea if divergence does not fall with intermediate feedback, if the stability boundary does not shift systematically with \(\alpha\), or if the recurrent module offers no benefit over simply increasing the static penalty.

## Disclaimer

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