Mesh-Free Physics-Informed PointNet (PIPN)

• Mesh-Free Physics-Informed PointNet (PIPN) is a deep learning framework that processes unordered point clouds with physics constraints from PDE residuals, boundary conditions, and sparse data.
• It couples a PointNet-style encoder with a per-point decoder to address multi-geometry problems in steady-state flow, thermal, porous media, and electromagnetic applications during a single training run.
• Key applications include fluid dynamics, inverse thermoelasticity, and electromagnetic scattering, showcasing its capability for generalizing across complex and irregular domains.

Mesh-Free Physics-Informed PointNet (PIPN) denotes a class of physics-informed learning methods in which a PointNet-style set encoder is coupled to PDE-constrained training on unstructured point clouds, so that the predicted field at a point depends on both its coordinates and a global representation of the computational domain. The formulation emerged to solve steady-state incompressible flow and thermal problems on multiple irregular geometries within a single training run, and it has since been instantiated for pore-scale Stokes flow in porous media, multi-geometry inverse thermoelasticity, biomechanics-constrained medical registration, coupled free-flow/porous-flow systems, KAN-based variants, and electromagnetic scattering on metasurface point clouds (Kashefi et al., 2022, Kashefi et al., 2022, Kashefi et al., 2023, Min et al., 2023, Ciceri et al., 15 Feb 2026, Kashefi et al., 8 Apr 2025, Armbruster et al., 30 Jul 2025).

1. Definition and conceptual scope

PIPN is a mesh-free, physics-informed deep learning framework whose input is an unordered point cloud representation of a domain and whose loss is built from PDE residuals, boundary conditions, and, when available, sparse observations. In the formulation introduced for steady incompressible flow on multiple irregular geometries, the solution is written pointwise as

$$\mathbf{y}_j = f\big(\mathbf{x}_j,\,g(\mathcal{X}_i)\big),$$

where $\mathcal{X}_i=\{\mathbf{x}_j\}_{j=1}^N$ is the point cloud of domain $V_i$, $g(\mathcal{X}_i)$ is a global latent feature extracted by a PointNet-style encoder, and $f$ is a per-point decoder (Kashefi et al., 2022). This explicitly makes the PDE solution a function of geometry, which is the principal departure from standard PINNs trained on a single fixed domain.

The framework is also positioned between classical PINNs and fully supervised multi-geometry surrogates. In the linear thermoelastic inverse problem on irregular domains, PIPN is described as filling the gap between PINNs, which are usually trained on one domain, and supervised models trained on thousands of labeled geometries: it solves an inverse linear elasticity problem on 532 irregular geometries simultaneously while using sparse labeled data per geometry (Kashefi et al., 2023). This suggests that PIPN is best understood not as one specific network topology but as a family of geometry-conditioned, point-cloud-based scientific machine learning models.

A frequent misconception is to equate “mesh-free” with “physics-free” or “boundary-free.” In the PIPN literature, mesh-free means that the neural solver does not use mesh connectivity, finite-difference stencils, or structured voxel grids; it does not mean that collocation points, boundary points, or observation points disappear. The point clouds still partition into interior, boundary, and sensor subsets, and PDE, BC, and data losses are evaluated on those subsets (Kashefi et al., 2022, Kashefi et al., 2022).

2. Architectural formulation

The canonical PIPN architecture follows the segmentation branch of PointNet. A shared pointwise encoder first maps each input coordinate to local features, a symmetric pooling operator extracts a permutation-invariant global descriptor, and a decoder concatenates the global descriptor back to per-point features to predict fields at every point. In the original multi-geometry incompressible-flow formulation, the encoder uses shared MLP blocks $(64,64)$ and $(64,64,1024)$, max pooling produces a 1024-dimensional global feature, and the decoder applies shared MLPs $(512,256,128)$ and $(128,n_{\text{PDE}})$ to output $(u,v,p)$ or $\mathcal{X}_i=\{\mathbf{x}_j\}_{j=1}^N$0 per point (Kashefi et al., 2022).

Later variants made the geometry encoding more explicit. For flows through and around porous structures, a geometry-aware PIPN augments each point with a signed distance function, a binary porous-region indicator, and one-hot boundary IDs, then concatenates these channels with early point features before global pooling (Ciceri et al., 15 Feb 2026). For pore-scale Stokes flow in porous media, the input is exclusively the pore space rather than both pore and grain space, the boundary is represented smoothly by the point cloud, and spatial resolution can vary across the domain (Kashefi et al., 2022). In electromagnetic scattering from metasurfaces, point features include coordinates and local permittivity $\mathcal{X}_i=\{\mathbf{x}_j\}_{j=1}^N$1, allowing the network to encode spatially varying material distributions directly in the point cloud representation (Armbruster et al., 30 Jul 2025).

Architectural specialization has become a major theme. “Physics-informed KAN PointNet” replaces shared MLPs with KAN layers parameterized by Jacobi polynomials and reports that degree $\mathcal{X}_i=\{\mathbf{x}_j\}_{j=1}^N$2 with $\mathcal{X}_i=\{\mathbf{x}_j\}_{j=1}^N$3 provides the best balance of accuracy and stability in the natural-convection inverse problem (Kashefi et al., 8 Apr 2025). In a different direction, PointNet is used as a permutation-invariant anatomical feature extractor in a biomechanics-constrained registration framework, where global features from source and target point sets are concatenated with source coordinates and passed through pointwise heads that predict displacement and stress (Min et al., 2023). These developments indicate that the PointNet component may serve either as a pure geometry encoder or as a broader set-conditioned latent representation for multiple physical fields.

3. Physics-informed objectives and mesh-free differentiation

The defining feature of PIPN is not PointNet alone but the way physical law is embedded in the loss. For incompressible flow, the loss is typically a weighted sum of continuity, momentum, boundary, and sparse-observation residuals. In the original PIPN for multiple irregular geometries, continuity and Navier–Stokes residuals are evaluated at interior points, velocity and pressure Dirichlet residuals at boundary points, and sparse observation losses at sensor locations, with all derivatives obtained by automatic differentiation with respect to the input coordinates (Kashefi et al., 2022). The porous-flow extension uses the same mechanism but adds the Darcy–Forchheimer drag term through a binary porous indicator $\mathcal{X}_i=\{\mathbf{x}_j\}_{j=1}^N$4, so that the residual reduces to standard Navier–Stokes in the free-flow region and includes porous resistance only inside $\mathcal{X}_i=\{\mathbf{x}_j\}_{j=1}^N$5 (Ciceri et al., 15 Feb 2026).

The same pattern appears outside fluid mechanics. In pore-scale porous-media flow, the loss combines the Stokes residuals, wall no-slip penalties, and sparse observations of velocity and pressure, again with first and second spatial derivatives computed by AD on the coordinate-to-field network (Kashefi et al., 2022). In multi-geometry inverse thermoelasticity, the loss combines the plane-stress thermoelastic PDE and sparse displacement sensors, enabling field completion over 532 domains using the same weights (Kashefi et al., 2023). In physics-informed medical image registration, the loss combines a Chamfer registration term with equilibrium, constitutive, and elastic-energy penalties for a 3D linear elastic model, while displacement gradients and stress divergence are computed by differentiating network outputs with respect to point coordinates (Min et al., 2023).

Some works generalize the physics core beyond standard strong-form residual minimization. “Physics-informed machine learning for reconstruction of dynamical systems with invariant measure score matching” develops a mesh-free PINN-IMSM framework in which the score of an invariant density is learned from unordered point samples and then embedded into a score-based steady Fokker–Planck constraint to reconstruct a drift field via minimal-energy PDE-constrained optimization (Chen et al., 19 Jan 2026). Although the implementation there uses pointwise MLPs rather than PointNet, the paper explicitly states that the conceptual machinery transfers to a PointNet-style architecture operating on unordered point sets. A plausible implication is that PIPN can be interpreted broadly as a geometry- or measure-conditioned point-cloud interface around an architecture-agnostic physics loss.

4. Application domains

PIPN has developed as a cross-domain methodology rather than a single-application tool. In steady incompressible flow, it was first demonstrated on multiple irregular 2D domains with cavity obstacles and later on thermally driven natural convection where the inner-cylinder temperature boundary condition is inferred from sparse observations (Kashefi et al., 2022). In porous media, it was adapted to pore-scale steady Stokes flow using only pore-space points and sparse point observations, with permeability recovered from the reconstructed velocity field (Kashefi et al., 2022). For coupled free-flow and porous-flow systems, geometry-aware PIPN enforces steady incompressible Navier–Stokes outside a porous body and a Darcy–Forchheimer extension inside it, with OpenFOAM-generated data in 2D ducts and 3D windbreak scenarios (Ciceri et al., 15 Feb 2026).

Solid mechanics applications are equally prominent. One PIPN formulation solves an inverse plane-stress thermoelastic problem on more than 500 irregular geometries simultaneously (Kashefi et al., 2023). Another uses PointNet within a biomechanics-constrained PINN for non-rigid MR–TRUS prostate registration, predicting displacements and stresses while enforcing linear elasticity on source-point collocation sets (Min et al., 2023). Related point-cloud operator models such as Point-DeepONet show that PointNet-based geometry encoding and SDF-conditioned trunks can accurately emulate nonlinear 3D structural response on non-parametric geometries, which suggests a natural pathway toward physics-informed operator variants even though that specific model is trained only with supervised FEM data (Park et al., 2024).

Wave problems have also adopted the PIPN template. For all-dielectric metasurfaces with inclined nanopillars, a mesh-free PIPN maps coordinate-plus-permittivity point clouds to complex electromagnetic fields constrained by the Helmholtz equation and periodic boundary conditions (Armbruster et al., 30 Jul 2025). The architecture generalizes across refractive-index combinations and pillar inclinations, which are treated as manufacturing defects. This breadth of applications indicates that PIPN is less tied to a specific PDE family than to a combination of three ingredients: point-cloud geometry representation, permutation-invariant encoding, and residual-based physical training.

5. Generalization, accuracy, and computational behavior

The empirical literature consistently treats multi-geometry generalization as a primary criterion. In the original incompressible-flow study, a single PIPN trained on 26 irregular geometries achieved average relative $\mathcal{X}_i=\{\mathbf{x}_j\}_{j=1}^N$6 errors of approximately $\mathcal{X}_i=\{\mathbf{x}_j\}_{j=1}^N$7 for $\mathcal{X}_i=\{\mathbf{x}_j\}_{j=1}^N$8, $\mathcal{X}_i=\{\mathbf{x}_j\}_{j=1}^N$9 for $V_i$0, and $V_i$1 for $V_i$2, while requiring 19,141 iterations and 51,673 s; the corresponding regular PINN baseline required 2,080,133 iterations and 902,361 s across 26 separate trainings (Kashefi et al., 2022). In the plane-stress thermoelastic inverse problem on 532 geometries, a well-tuned PIPN achieved average relative $V_i$3 errors of about $V_i$4 and $V_i$5 for the two displacement components (Kashefi et al., 2023). These results do not establish universal superiority over classical discretizations, but they do show that point-cloud amortization substantially alters the computational trade-off relative to single-domain PINNs.

In porous media, the pore-scale Stokes PIPN obtained velocity errors on the order of $V_i$6–$V_i$7, pressure errors between roughly $V_i$8 and $V_i$9, and permeability errors below $g(\mathcal{X}_i)$0 across three correlation-length regimes (Kashefi et al., 2022). In coupled free-flow/porous-flow simulations, geometry-aware PIPN on 2D ducts with fixed boundary conditions reached test-set MAEs of $g(\mathcal{X}_i)$1 for $g(\mathcal{X}_i)$2, $g(\mathcal{X}_i)$3 for $g(\mathcal{X}_i)$4, and $g(\mathcal{X}_i)$5 for $g(\mathcal{X}_i)$6, with unseen-geometry errors remaining in the $g(\mathcal{X}_i)$7–$g(\mathcal{X}_i)$8 range; inference was about $g(\mathcal{X}_i)$9 faster than OpenFOAM for that case (Ciceri et al., 15 Feb 2026). The same study used PI-GANO rather than PIPN for the 3D windbreak setting, which is a reminder that point-cloud PIPN is not the only route to geometry generalization.

In electromagnetic scattering, the metasurface PIPN reached a near-field MAPE of $f$0 and diffraction-efficiency APE of $f$1 for a 2D SiO$f$2 case, $f$3 near-field MAPE in 3D, and $f$4 near-field MAPE for the higher-contrast TiO$f$5-on-SiO$f$6 case (Armbruster et al., 30 Jul 2025). In medical registration, the multi-patient PointNet-PINN model reduced test TRE from $f$7 mm to $f$8 mm while improving biomechanical plausibility as measured by deformation-magnitude ratios between rigid and soft regions (Min et al., 2023). The KAN-based variant reports that, with approximately the same number of trainable parameters and comparable epoch cost, PI-KAN-PointNet yields more accurate predictions than MLP-based PIPN, particularly for unknown boundary values on nonsmooth geometries (Kashefi et al., 8 Apr 2025).

6. Limitations, misconceptions, and adjacent directions

Reported failure modes are remarkably consistent across domains. Errors concentrate near sharp corners, material interfaces, porous–fluid transitions, wakes, and other large-gradient regions (Kashefi et al., 2022, Ciceri et al., 15 Feb 2026, Armbruster et al., 30 Jul 2025). In porous-media flow, omitting pressure observations leaves velocity predictions largely intact but causes the pressure field itself to become inaccurate, because the momentum equations constrain pressure gradients more strongly than absolute pressure (Kashefi et al., 2022). In the porous-structure flow study, performance degrades mainly near sharp interfaces and in regions with large gradients, and the models are limited to laminar regimes (Ciceri et al., 15 Feb 2026). In metasurface scattering, errors increase for high refractive-index contrast and near material interfaces (Armbruster et al., 30 Jul 2025). These are not incidental implementation artifacts; they reflect a structural tension between global set encoders and localized singular or near-singular physics.

A second misconception is that PointNet alone solves the multi-geometry PINN problem. The literature shows that geometry encoding is necessary but not sufficient. Loss construction, collocation strategy, derivative quality, and observation design all remain decisive. This is evident in studies that compare conservative and non-conservative momentum forms, include or remove pressure observations, vary sensor spacing, or alter point-cloud density (Kashefi et al., 2022, Kashefi et al., 2022). It is also evident in neighboring mesh-free point-based directions. GPINNs emphasize global nodal association rather than PointNet-style geometry encoding (Chen et al., 9 Mar 2025); WE-PINNs replace strong-form residual minimization with weak and entropy control-volume constraints for conservation laws with shocks (Oubarka et al., 25 Mar 2026); Point-DeepONet and PI-GANO shift toward operator learning on point clouds and geometry-conditioned latent spaces (Park et al., 2024, Ciceri et al., 15 Feb 2026).

Current extensions suggest two plausible trajectories for PIPN. One is architectural: PointNet++, graph neural networks, KPConv, KAN layers, and other local or spectral point operators can replace the shared MLP blocks while keeping the same physics-informed point-cloud training logic (Kashefi et al., 8 Apr 2025, Ciceri et al., 15 Feb 2026). The other is physical: the same mesh-free set-processing backbone can host score-based Fokker–Planck reconstruction, weak conservation-law formulations, elasticity with constitutive constraints, or wave scattering with material-conditioned point features (Chen et al., 19 Jan 2026, Min et al., 2023, Armbruster et al., 30 Jul 2025). This suggests that “Mesh-Free Physics-Informed PointNet” is most accurately treated as a methodological family: a point-cloud, permutation-invariant, geometry-aware interface through which diverse physics-informed objectives can be imposed on irregular domains without mesh connectivity.