Adaptive Feature Capture Method (AFCM) in PDEs

• Adaptive Feature Capture Method (AFCM) is a mesh-free machine learning approach for solving PDEs with low-regularity solutions by concentrating approximation power in steep-gradient regions.
• It extends the Random Feature Method by regenerating feature parameters and hyperplane geometry based on gradient norms, thereby enhancing local resolution without increasing overall model size.
• AFCM couples monitor-driven basis reinitialization with adaptive collocation redistribution, offering improved accuracy on challenging PDE problems while preserving computational efficiency.

Searching arXiv for the most relevant AFCM and adjacent adaptive feature-capture papers. arXiv search query: "Adaptive Feature Capture Method low regularity PDEs random feature method" Adaptive Feature Capture Method (AFCM) denotes a mesh-free machine-learning framework proposed for partial differential equations with low-regularity solutions, where accuracy is limited not primarily by model size but by the inability to concentrate approximation power in steep-gradient or near-singular regions. In its canonical formulation, AFCM extends the Random Feature Method (RFM) by adaptively redistributing both random-feature geometry and collocation points according to the gradient norm of an approximate solution, thereby clustering partition hyperplanes and residual-enforcement points where local resolution is most needed without increasing their total number (Deng et al., 17 Jul 2025).

1. Problem class and conceptual position

AFCM is motivated by PDEs whose solutions exhibit sharply localized peaks, near-singular layers, line singularities, steep transition fronts, moving localized structures in time-dependent problems, or strong convection/diffusion boundary-layer-type behavior. The core model problem is a boundary value problem

$$\begin{cases} \mathcal{L}\boldsymbol{\phi}(\boldsymbol{x})=\boldsymbol{f}(\boldsymbol{x}), & \boldsymbol{x}\in \Omega,\ \mathcal{B}\boldsymbol{\phi}(\boldsymbol{x})=\boldsymbol{g}(\boldsymbol{x}), & \boldsymbol{x}\in \partial\Omega, \end{cases}$$

with nonlinear PDEs handled through iterative linearization such as Picard iteration (Deng et al., 17 Jul 2025).

The method is positioned between adaptive mesh methods and randomized neural approximations. Traditional mesh-based solvers can resolve low-regularity structure through adaptive mesh refinement or moving meshes, but mesh generation and adaptive refinement become expensive in complex geometries. PINNs and related deep-learning PDE solvers are mesh-free, yet are described as lacking reliable error control and as not naturally concentrating approximation power where the solution is hardest to represent. Standard RFM is efficient because only a linear least-squares problem is solved after random feature construction, but it distributes feature partition hyperplanes and collocation points essentially uniformly, which limits resolution in steep local structures (Deng et al., 17 Jul 2025).

A common misconception is to treat AFCM as a moving-mesh algorithm in disguise. It is not. AFCM does not move a mesh; it transfers the moving-mesh idea into random-feature space. Mesh density is replaced by partition-hyperplane density, and mesh concentration is replaced by adaptive redistribution of feature parameters and collocation points. Another misconception is to view AFCM as a fixed-basis reweighting scheme. In fact, the feature parameters are reinitialized during adaptation, so the approximation space itself is reconfigured rather than merely reweighted (Deng et al., 17 Jul 2025).

2. Approximation architecture inherited from RFM

AFCM starts from the RFM decomposition of the domain into $M_p$ non-overlapping subdomains,

$\Omega=\bigcup_{n=1}^{M_p}\Omega_n,$

with local normalization

$$\tilde{\boldsymbol{x}}=\frac{1}{\boldsymbol{r}_n}\left(\boldsymbol{x}-\boldsymbol{x}_n\right), \qquad n=1,\dots,M_p.$$

A partition of unity $\psi_n$ is associated with each $\Omega_n$. In the AFCM experiments, the indicator-type PoU

$$\psi_n(\boldsymbol{x})=\mathbb{I}_{[-1,1]}(\tilde{\boldsymbol{x}})$$

is used, while inter-subdomain continuity is enforced by regularization with $C^1$ continuity following Dong and Li (Deng et al., 17 Jul 2025).

Local random features are written as

$$\varphi_{nj}(\boldsymbol{x})=\sigma\!\left(\boldsymbol{W}_{nj}\cdot \tilde{\boldsymbol{x}}+b_{nj}\right),$$

and the global approximation takes the PoU-coupled form

$$\tilde{\boldsymbol{\phi}}(\boldsymbol{x}) = \left( \sum_{n=1}^{M_p}\psi_n(\boldsymbol{x})\sum_{j=1}^{J_n}u_{nj}^1 \varphi_{nj}^1(\boldsymbol{x}), \dots, \sum_{n=1}^{M_p}\psi_n(\boldsymbol{x})\sum_{j=1}^{J_n}u_{nj}^{d_\phi}\varphi_{nj}^{d_\phi}(\boldsymbol{x}) \right)^T.$$

The unknowns are the output coefficients $M_p$0, so the solve remains linear in the trainable variables (Deng et al., 17 Jul 2025).

Residual minimization is performed by collocation. For collocation points $M_p$1, the least-squares loss is

$M_p$2

which yields a linear least-squares system $M_p$3. The diagonal rescaling matrices $M_p$4 and $M_p$5 normalize operator magnitudes through the maxima of $M_p$6 and $M_p$7, respectively (Deng et al., 17 Jul 2025).

AFCM adopts the TransNet parameterization

$M_p$8

where $M_p$9 is a unit normal, $\Omega=\bigcup_{n=1}^{M_p}\Omega_n,$0 is a hyperplane offset, and $\Omega=\bigcup_{n=1}^{M_p}\Omega_n,$1 controls steepness. The induced partition hyperplane is

$\Omega=\bigcup_{n=1}^{M_p}\Omega_n,$2

Its local density is measured by

$\Omega=\bigcup_{n=1}^{M_p}\Omega_n,$3

with

$\Omega=\bigcup_{n=1}^{M_p}\Omega_n,$4

This density becomes the geometric object that AFCM adapts (Deng et al., 17 Jul 2025).

3. Adaptive feature capture mechanism

The defining AFCM step is to convert the gradient norm of the current approximation into a monitor function. A uniformly sampled set $\Omega=\bigcup_{n=1}^{M_p}\Omega_n,$5 of $\Omega=\bigcup_{n=1}^{M_p}\Omega_n,$6 points in $\Omega=\bigcup_{n=1}^{M_p}\Omega_n,$7 is generated, and a discrete probability density is defined by

$\Omega=\bigcup_{n=1}^{M_p}\Omega_n,$8

where $\Omega=\bigcup_{n=1}^{M_p}\Omega_n,$9 prevents pathological over-concentration. Regions with larger $$\tilde{\boldsymbol{x}}=\frac{1}{\boldsymbol{r}_n}\left(\boldsymbol{x}-\boldsymbol{x}_n\right), \qquad n=1,\dots,M_p.$$0 therefore receive higher sampling probability (Deng et al., 17 Jul 2025).

Feature adaptation uses weighted random sampling from $$\tilde{\boldsymbol{x}}=\frac{1}{\boldsymbol{r}_n}\left(\boldsymbol{x}-\boldsymbol{x}_n\right), \qquad n=1,\dots,M_p.$$1 according to $$\tilde{\boldsymbol{x}}=\frac{1}{\boldsymbol{r}_n}\left(\boldsymbol{x}-\boldsymbol{x}_n\right), \qquad n=1,\dots,M_p.$$2. If $$\tilde{\boldsymbol{x}}=\frac{1}{\boldsymbol{r}_n}\left(\boldsymbol{x}-\boldsymbol{x}_n\right), \qquad n=1,\dots,M_p.$$3 are the selected points in subdomain $$\tilde{\boldsymbol{x}}=\frac{1}{\boldsymbol{r}_n}\left(\boldsymbol{x}-\boldsymbol{x}_n\right), \qquad n=1,\dots,M_p.$$4, the feature steepness parameters are reset by

$$\tilde{\boldsymbol{x}}=\frac{1}{\boldsymbol{r}_n}\left(\boldsymbol{x}-\boldsymbol{x}_n\right), \qquad n=1,\dots,M_p.$$5

where $$\tilde{\boldsymbol{x}}=\frac{1}{\boldsymbol{r}_n}\left(\boldsymbol{x}-\boldsymbol{x}_n\right), \qquad n=1,\dots,M_p.$$6 limits excessive variation. Hyperplane geometry is then reinitialized through

$$\tilde{\boldsymbol{x}}=\frac{1}{\boldsymbol{r}_n}\left(\boldsymbol{x}-\boldsymbol{x}_n\right), \qquad n=1,\dots,M_p.$$7

This construction forces the hyperplane

$$\tilde{\boldsymbol{x}}=\frac{1}{\boldsymbol{r}_n}\left(\boldsymbol{x}-\boldsymbol{x}_n\right), \qquad n=1,\dots,M_p.$$8

to pass through $$\tilde{\boldsymbol{x}}=\frac{1}{\boldsymbol{r}_n}\left(\boldsymbol{x}-\boldsymbol{x}_n\right), \qquad n=1,\dots,M_p.$$9. High-gradient sample points therefore directly attract hyperplanes and, via larger $\psi_n$0, sharper local nonlinearities (Deng et al., 17 Jul 2025).

Collocation adaptation uses the same density. Only interior collocation points are resampled; boundary and interface points are retained. If $\psi_n$1 denotes the newly sampled interior set in $\psi_n$2, the collocation set becomes

$\psi_n$3

The approximation budget is therefore reallocated in two coupled ways: the basis functions become denser and steeper near difficult regions, and the residual is enforced more strongly there (Deng et al., 17 Jul 2025).

This mechanism is adaptive in a stricter sense than feature weighting. The random features do not remain fixed. They are regenerated repeatedly through new anchor points, new offsets, and new shape parameters. A plausible implication is that AFCM changes both approximation geometry and sampling geometry simultaneously, which is why it can increase local expressive power without increasing the total number of features or collocation points.

4. Algorithmic workflow and implementation profile

The operational pipeline is iterative. The domain is first partitioned, initial collocation points are sampled in each subdomain, and initial TransNet parameters are generated using

$\psi_n$4

with $\psi_n$5 drawn from a standard Gaussian and $\psi_n$6. A common subdomain-level shape parameter $\psi_n$7 is selected by Gaussian random field fitting via the loss

$\psi_n$8

followed by

$\psi_n$9

An initial least-squares solve then yields $\Omega_n$0 (Deng et al., 17 Jul 2025).

The adaptive loop proceeds by computing the monitor density on a uniformly sampled pool $\Omega_n$1, resampling feature anchor points and interior collocation points according to $\Omega_n$2, reconstructing the basis, and resolving the least-squares system. The pseudocode uses a fixed number $\Omega_n$3 of adaptation steps, although the prose also states that iteration may continue until convergence. For nonlinear PDEs, the same linear least-squares machinery is embedded inside iterative linearization rather than replaced by full hidden-parameter optimization (Deng et al., 17 Jul 2025).

The reported common experimental settings are concrete. Unless otherwise stated, the experiments use $\Omega_n$4, $\Omega_n$5 collocation points per subdomain, $\Omega_n$6 interior points per subdomain, $\Omega_n$7, $\Omega_n$8 Gaussian random field realizations, $\Omega_n$9, $$\psi_n(\boldsymbol{x})=\mathbb{I}_{[-1,1]}(\tilde{\boldsymbol{x}})$$0, density bandwidth $$\psi_n(\boldsymbol{x})=\mathbb{I}_{[-1,1]}(\tilde{\boldsymbol{x}})$$1, correlation length $$\psi_n(\boldsymbol{x})=\mathbb{I}_{[-1,1]}(\tilde{\boldsymbol{x}})$$2, activation $$\psi_n(\boldsymbol{x})=\mathbb{I}_{[-1,1]}(\tilde{\boldsymbol{x}})$$3, and torch.linalg.lstsq for the least-squares solve (Deng et al., 17 Jul 2025).

Several formal properties remain open. The paper does not provide a convergence theorem for AFCM, a conditioning analysis of the adaptive least-squares matrix, a complexity theorem, or a rigorous a posteriori error estimator beyond the gradient-based monitor function. These omissions matter because the method is algorithmically clear and empirically motivated, but its theoretical status remains incomplete (Deng et al., 17 Jul 2025).

5. Empirical behavior, error metrics, and known limitations

AFCM is evaluated with relative $$\psi_n(\boldsymbol{x})=\mathbb{I}_{[-1,1]}(\tilde{\boldsymbol{x}})$$4 and $$\psi_n(\boldsymbol{x})=\mathbb{I}_{[-1,1]}(\tilde{\boldsymbol{x}})$$5 errors,

$$\psi_n(\boldsymbol{x})=\mathbb{I}_{[-1,1]}(\tilde{\boldsymbol{x}})$$6

The abstract-level conclusion is that numerical experiments demonstrate effectiveness in accurately resolving low-regularity problems, including settings with complex geometries, while preserving the mesh-free efficiency of RFM (Deng et al., 17 Jul 2025).

One documented benchmark is a two-dimensional Poisson problem on $$\psi_n(\boldsymbol{x})=\mathbb{I}_{[-1,1]}(\tilde{\boldsymbol{x}})$$7 with exact solution

$$\psi_n(\boldsymbol{x})=\mathbb{I}_{[-1,1]}(\tilde{\boldsymbol{x}})$$8

which exhibits a sharply localized peak at the origin. The monitor-sampling pool size is $$\psi_n(\boldsymbol{x})=\mathbb{I}_{[-1,1]}(\tilde{\boldsymbol{x}})$$9, and experiments are reported for $C^1$0, with initial shape parameters $C^1$1, respectively. Even from this partially reported setup, the intended role of AFCM is unambiguous: it reallocates feature hyperplanes and collocation points toward the singular peak rather than increasing global model size (Deng et al., 17 Jul 2025).

The principal empirical claim is therefore qualitative but specific: AFCM improves local resolution by redistributing neurons and collocation points in high-gradient regions. A plausible implication is that its gains come from targeted concentration of basis geometry rather than from overparameterization. This distinguishes AFCM from simply enlarging the random-feature dictionary.

Its limitations are equally clear. The method relies on the gradient norm of an approximate solution as the monitor, so adaptation quality depends on the quality of the current approximation. The constants $C^1$2 and $C^1$3 regularize the redistribution but are heuristic. The stopping rule is iteration-count based in the pseudocode. The absence of convergence and complexity theory means AFCM is presently best regarded as a robust adaptive computational framework rather than a fully characterized numerical method (Deng et al., 17 Jul 2025).

6. Broader usage of “adaptive feature capture” and adjacent paradigms

In the strict sense, AFCM refers to the PDE-oriented method just described (Deng et al., 17 Jul 2025). In a broader functional sense, related ideas appear across several research areas under different names. Small-object detection in DETR variants uses fine-grained path augmentation and adaptive feature fusion to preserve low-level detail and adaptively combine multi-scale features before decoding, addressing the combination of missing fine detail and insufficiently adaptive multi-scale fusion (Huang et al., 2024). Sparse 3D semantic segmentation uses an encoder-side attentive feature fusion module and a decoder-side adaptive feature selection module to weight branch outputs across receptive fields and re-weight multi-scale skip features (Cheng et al., 2021). Diffusion-model ensembling uses blockwise spatially aware aggregation of intermediate U-Net features, with per-model attention maps conditioned on prompt, timestep, latent features, and spatial position (Wang et al., 2024). Wireless foundation models use routing-based adaptive composition of hidden states from several Transformer depths, replacing final-layer-only extraction with task- and sample-conditional weighted summation of low-, mid-, and high-level features (Shi et al., 9 Jun 2026).

These works are not instances of AFCM in the exact nomenclature of the PDE paper. Their commonality is structural rather than terminological: each replaces a fixed feature path with an input- or task-adaptive redistribution of representational emphasis. In that broader taxonomy, AFCM for PDEs is the version in which adaptive emphasis is realized geometrically, through hyperplane placement and collocation density, instead of through attention, gating, or learned fusion weights.

This broader context also clarifies what AFCM is not. It is not merely adaptive feature fusion, because it changes the geometry of the feature basis itself. It is not only adaptive sampling, because it simultaneously redistributes basis functions. It is not adaptive mesh refinement, because no mesh is evolved. Its distinctive identity lies in coupling monitor-driven basis reinitialization with monitor-driven collocation redistribution inside a random-feature least-squares solver (Deng et al., 17 Jul 2025).