Inverse-Parameter Basis PINNs

• The paper introduces IP-Basis PINNs as inverse formulations where unknown states or parameters are approximated via explicit or learned basis expansions.
• They employ offline-online meta-learning and spectral meshless methods to enable rapid multi-query inference and address inverse problems on irregular domains.
• The approach integrates PINN-style residual losses with analytical basis derivatives, achieving exponential convergence and improved conditioning in various benchmarks.

Searching arXiv for the cited papers and closely related IP-Basis PINN work. (Roggeveen et al., 29 Oct 2025) IP-Basis PINN spectral basis irregular geometry

Inverse-Parameter Basis PINNs (IP-Basis PINNs) are inverse physics-informed formulations in which the unknown state, the unknown parameters, or both are represented through an explicit basis while PDE residuals, boundary and initial conditions, observational data, and optional optimization targets are enforced through a PINN-style loss. In the current arXiv literature, the term is used explicitly for two distinct but related constructions: an offline-online meta-learning framework that pretrains a shared basis for rapid multi-query inverse inference, and a spectral meshless formulation that replaces the neural network itself with a compact tensor-product basis on a containing hyperrectangle for inverse PDEs on irregular geometries (Manor et al., 8 Sep 2025, Roggeveen et al., 29 Oct 2025). Closely related basis-centric inverse PINN designs include hybrid PINN–IGA formulations for source recovery, RBF-based kernel identification in peridynamics, and parameter-basis representations for PBPK brain models (Mardal et al., 2023, Difonzo et al., 2023, Wickramasinghe et al., 16 Sep 2025).

1. Definition and problem class

The common problem class is an inverse differential system in which one seeks a state $u$ and unknown parameters $p$ so that a differential operator is satisfied together with geometric and data constraints. In the irregular-geometry spectral formulation, the inverse problem is posed on a possibly irregular domain $\Omega \subset \mathbb{R}^d$, with

$$F(u,\nabla u,\nabla^2 u,\ldots,p;x,t)=0,\qquad (x,t)\in \Omega\times[t_0,t_1],$$

subject to boundary and initial conditions, where $p$ may include material coefficients, source terms, and boundary inputs (Roggeveen et al., 29 Oct 2025). In the multi-query formulation, one instead considers a parametric class of PDEs or ODEs, with many observed datasets $D_q$ generated by different true parameters $\theta_q^\*$, and aims to amortize inverse inference across queries by learning a shared solution basis offline and only fitting a lightweight online head and the inverse parameters per query (Manor et al., 8 Sep 2025).

What makes these methods “basis” methods is not a single canonical parameterization but the replacement of a fully free state or parameter representation by a structured finite or learned expansion. In one line of work, the state is written directly as

$u(x,t)=\sum_{k=1}^{K} c_k \phi_k(x,t),$

or, in tensor form,

$u(x)=C:\Phi(x),$

with $\Phi(x)=\phi_1(x_1)\otimes\cdots\otimes\phi_d(x_d)$ on a hyperrectangle $p$0 containing $p$1 (Roggeveen et al., 29 Oct 2025). In another line, a frozen trunk $p$2 provides $p$3 features $p$4, and the online inverse problem is solved through a linear readout $p$5, so that $p$6 while $p$7 is optimized jointly (Manor et al., 8 Sep 2025). Related formulations parameterize only the inverse field, for example $p$8 in an IGA basis while the state $p$9 remains a PINN (Mardal et al., 2023).

This usage establishes IP-Basis PINNs as a family of inverse formulations rather than a single architecture. A common misconception is that the term refers only to a meta-learning surrogate or only to a spectral discretization. The present literature instead supports both usages explicitly, with the unifying idea that inverse estimation is stabilized or accelerated by moving from generic function approximation to a basis-adapted parameterization (Manor et al., 8 Sep 2025, Roggeveen et al., 29 Oct 2025).

2. Principal architectural realizations

Two explicit realizations dominate the literature. The first is the offline-online IP-Basis PINN. Here a deep network is decomposed into a shared trunk $\Omega \subset \mathbb{R}^d$0 and a linear output head $\Omega \subset \mathbb{R}^d$1. Offline, many readouts are optimized in parallel for different parameter samples, and the trunk is trained so that its outputs span the solution manifold across the parameter distribution. Online, the trunk is frozen and only a fresh linear head and the inverse parameters are optimized against the new observations and the PDE loss. This design targets the multi-query setting in which standard PINNs would require retraining from scratch for every dataset (Manor et al., 8 Sep 2025).

The second is the spectral meshless IP-Basis PINN. Here the neural network is removed entirely from the state representation and replaced by global tensor-product bases such as Chebyshev, Legendre, Fourier, or mixed bases on a containing hyperrectangle $\Omega \subset \mathbb{R}^d$2. Irregular geometry is handled by restricting collocation to $\Omega \subset \mathbb{R}^d$3 and $\Omega \subset \mathbb{R}^d$4, so the basis construction is separated from the geometry while retaining a PINN-style residual minimization framework (Roggeveen et al., 29 Oct 2025).

A broader basis-centric landscape is visible in related inverse PINN work. The hybrid IGA–PINN formulation uses tensor-product B-splines or NURBS for the unknown source field and a PINN for the state. The RBF-iPINN for peridynamics uses a radial basis function at the first layer of the kernel subnet to encode evenness and locality of the micromodulus $\Omega \subset \mathbb{R}^d$5. The PBPK-iPINN literature frames an IP-Basis extension as the representation of physiological parameters through a low-dimensional basis $\Omega \subset \mathbb{R}^d$6, possibly with compartmental or temporal structure (Mardal et al., 2023, Difonzo et al., 2023, Wickramasinghe et al., 16 Sep 2025).

Realization	Basis object	Inverse role
Offline-online IP-Basis PINN	Frozen trunk features with linear readout	Rapid multi-query solution and parameter inference
Spectral meshless IP-Basis PINN	Chebyshev, Legendre, Fourier, or mixed tensor-product bases on $\Omega \subset \mathbb{R}^d$7	Joint optimization of coefficients and unknown PDE parameters on irregular $\Omega \subset \mathbb{R}^d$8
Hybrid and related basis-centric variants	IGA bases, RBF input encodings, parameter bases $\Omega \subset \mathbb{R}^d$9	Source-field inversion, constitutive-kernel identification, or structured parameter estimation

This taxonomy suggests that the decisive design choice is where the basis is introduced: in the state, in the parameter field, in the network input encoding, or in a frozen latent representation. A plausible implication is that different IP-Basis PINN variants can be interpreted as choosing different coordinates for the inverse search space rather than altering the PINN principle itself.

3. Loss construction, inverse optimization, and autodiff

Across formulations, the optimization problem is built from a composite loss. In the spectral meshless setting, the objective over coefficients $$F(u,\nabla u,\nabla^2 u,\ldots,p;x,t)=0,\qquad (x,t)\in \Omega\times[t_0,t_1],$$0 and unknown parameters $$F(u,\nabla u,\nabla^2 u,\ldots,p;x,t)=0,\qquad (x,t)\in \Omega\times[t_0,t_1],$$1 is

$$F(u,\nabla u,\nabla^2 u,\ldots,p;x,t)=0,\qquad (x,t)\in \Omega\times[t_0,t_1],$$2

with support for Dirichlet, Neumann, Robin, and periodic boundary conditions, data assimilation, and optimization over task functionals $$F(u,\nabla u,\nabla^2 u,\ldots,p;x,t)=0,\qquad (x,t)\in \Omega\times[t_0,t_1],$$3 through $$F(u,\nabla u,\nabla^2 u,\ldots,p;x,t)=0,\qquad (x,t)\in \Omega\times[t_0,t_1],$$4 (Roggeveen et al., 29 Oct 2025). Unknown parameters are optimized jointly with the basis coefficients, and the method gives explicit expressions for $$F(u,\nabla u,\nabla^2 u,\ldots,p;x,t)=0,\qquad (x,t)\in \Omega\times[t_0,t_1],$$5 and $$F(u,\nabla u,\nabla^2 u,\ldots,p;x,t)=0,\qquad (x,t)\in \Omega\times[t_0,t_1],$$6, while using autodiff only at the level of the composite loss.

The offline-online framework uses a different decomposition of the same basic principle. Offline, the loss averages standard PINN losses over many readouts associated with sampled parameter instances. Online, the frozen basis features are combined through a trainable linear head, and the new inverse problem minimizes

$$F(u,\nabla u,\nabla^2 u,\ldots,p;x,t)=0,\qquad (x,t)\in \Omega\times[t_0,t_1],$$7

A non-trivial validation and early-stopping mechanism is introduced for offline training: a fresh validation head is optimized by a single gradient step per epoch, and checkpointing is based on the resulting validation loss (Manor et al., 8 Sep 2025).

The hybrid IGA inverse formulation provides a variational least-squares foundation for an inverse IP-Basis PINN with state $$F(u,\nabla u,\nabla^2 u,\ldots,p;x,t)=0,\qquad (x,t)\in \Omega\times[t_0,t_1],$$8 and source basis coefficients $$F(u,\nabla u,\nabla^2 u,\ldots,p;x,t)=0,\qquad (x,t)\in \Omega\times[t_0,t_1],$$9. Its loss is

$p$0

and the paper states that the inverse least-squares problem is well-posed if $p$1 and the solution $p$2 admits $p$3 regularity (Mardal et al., 2023).

Loss balancing is a recurrent technical issue. The spectral meshless method supports adaptive loss reweighting through

$p$4

the PBPK-iPINN study reports that convergence to the correct solution requires appropriate weighting of data, initial-condition, and residual losses, and the peridynamic RBF-iPINN adds an explicit symmetry loss $p$5 alongside PDE and data terms (Roggeveen et al., 29 Oct 2025, Wickramasinghe et al., 16 Sep 2025, Difonzo et al., 2023).

4. Basis design, geometry handling, and operator evaluation

The technical core of IP-Basis PINNs lies in the basis choice. In the spectral meshless method, each one-dimensional basis $p$6 may be chosen independently and anisotropically: Chebyshev polynomials $p$7, Legendre polynomials $p$8, Fourier series, or mixed bases, with time treated as an additional coordinate. This allows anisotropic truncations $p$9 to reflect stiffness or anisotropy, and periodic boundary conditions can be built in through Fourier modes rather than imposed only by penalties (Roggeveen et al., 29 Oct 2025).

Irregular geometries are handled by defining the basis on a containing hyperrectangle $D_q$0 and restricting residual evaluation to $D_q$1. The paper describes this as a Fourier-extension-style strategy: collocation points are sampled only inside $D_q$2 and on $D_q$3, or residuals are multiplied by a mask $D_q$4. Boundaries may be specified through a binary mask, a level set, or an analytic parameterization $D_q$5, and boundary normals may be obtained from an analytic boundary parameterization or from gradients of a signed distance or mask. Time-dependent spatiotemporal problems are then solved without time-stepping because time is just another coordinate in the basis (Roggeveen et al., 29 Oct 2025).

Differential operators are treated differently across variants, but always in a basis-aware way. In the spectral formulation, basis derivatives are analytic: $D_q$6 and tensor-product derivatives follow by replacing the relevant one-dimensional factor by its derivative. As a consequence, $D_q$7, $D_q$8, and higher derivatives can be precomputed so that residual assembly reduces to fast matrix multiplications. In the offline-online IP-Basis PINN, the trunk is frozen and derivatives of the basis features are assembled linearly, while forward-mode automatic differentiation with hyper-dual numbers is used to compute first and second derivatives of all outputs efficiently in one forward pass (Roggeveen et al., 29 Oct 2025, Manor et al., 8 Sep 2025).

Related basis choices serve different physical priors. In the RBF-iPINN, the first layer uses a radial basis function

$D_q$9

with inverse quadratic or multiquadric forms, to encode the evenness and locality of the peridynamic kernel. In the IGA-guided inverse formulation, the key design rule is the containment condition $\theta_q^\*$0, implemented in 2D through $\theta_q^\*$1 with $\theta_q^\*$2 and $\theta_q^\*$3. In the PBPK setting, the proposed IP-Basis extension writes global or time-varying physiological parameters as $\theta_q^\*$4 or $\theta_q^\*$5, thereby shifting the inverse task from direct parameter recovery to coefficient recovery on a structured basis (Difonzo et al., 2023, Mardal et al., 2023, Wickramasinghe et al., 16 Sep 2025).

5. Benchmarked applications and empirical behavior

The empirical record spans PDEs, ODEs, inverse field estimation, and control. The spectral meshless IP-Basis PINN is validated on Laplace problems on a peanut-with-hole domain and the Lake Taal boundary, the wave equation on an irregular domain, diffusion on the sphere surface, Allen–Cahn, nonlinear Schrödinger, shallow shelf approximation viscosity inversion, heat control, and optimal transport. The paper reports empirical exponential convergence across diverse problems, including exponential decay of $\theta_q^\*$6 and $\theta_q^\*$7 errors with basis size for Laplace on irregular domains, exponential convergence with anisotropic truncation for Allen–Cahn and nonlinear Schrödinger, and accurate inversion of Antarctic ice-shelf viscosity from synthetic and real Amery Ice Shelf data. It further states that the approach matches COMSOL and analytic references with exponential error decay in $\theta_q^\*$8, that runtime is minutes on a laptop CPU with far fewer parameters than PINNs, and that specialized forward solvers remain faster for pure forward problems while the method excels in inverse and design flexibility (Roggeveen et al., 29 Oct 2025).

The offline-online IP-Basis PINN is benchmarked on the damped harmonic oscillator, Lotka–Volterra with unknown functional terms, and the quantum harmonic oscillator. Reported results include 5–6× offline speedups from forward-mode autodiff, approximately $\theta_q^\*$9 speedup in the quantum harmonic oscillator multi-query test case relative to standard PINNs, and low parameter MSE for 10k and 1k data points with gradual degradation at 100 and 10 points; at 10 points the reported parameter MAE is approximately $u(x,t)=\sum_{k=1}^{K} c_k \phi_k(x,t),$0. The same study reports that standard PINNs often have lower error when data is plentiful, whereas the basis method remains usable under very scarce data (Manor et al., 8 Sep 2025).

The variational IGA analysis supplies a quantitative argument for parameter-basis design. In the inverse elliptic example, enforcing $u(x,t)=\sum_{k=1}^{K} c_k \phi_k(x,t),$1 yields nearly uniform $u(x,t)=\sum_{k=1}^{K} c_k \phi_k(x,t),$2 $u(x,t)=\sum_{k=1}^{K} c_k \phi_k(x,t),$3 error for $u(x,t)=\sum_{k=1}^{K} c_k \phi_k(x,t),$4, versus errors up to $u(x,t)=\sum_{k=1}^{K} c_k \phi_k(x,t),$5 without containment. The authors also observe notable error reduction in inverse neural-network experiments when representability analogous to containment is enforced through activation design (Mardal et al., 2023).

In the peridynamic setting, the RBF-iPINN recovers V-shaped and compactly supported kernels, and for the Gaussian kernel example the subsequent parametric inverse PINN reports recovered parameters $u(x,t)=\sum_{k=1}^{K} c_k \phi_k(x,t),$6 and $u(x,t)=\sum_{k=1}^{K} c_k \phi_k(x,t),$7, compared with exact values $u(x,t)=\sum_{k=1}^{K} c_k \phi_k(x,t),$8 and $u(x,t)=\sum_{k=1}^{K} c_k \phi_k(x,t),$9. In the PBPK brain model, the inverse PINN recovers compartment volumes with very small absolute errors, while $u(x)=C:\Phi(x),$0 and $u(x)=C:\Phi(x),$1 are more difficult; one reported run gives $u(x)=C:\Phi(x),$2 error $u(x)=C:\Phi(x),$3 and $u(x)=C:\Phi(x),$4 error $u(x)=C:\Phi(x),$5, and the best model after approximately $u(x)=C:\Phi(x),$6 iterations has total loss $u(x)=C:\Phi(x),$7 (Difonzo et al., 2023, Wickramasinghe et al., 16 Sep 2025).

Taken together, these benchmarks suggest a division of labor. IP-Basis PINNs are repeatedly strongest in inverse, control, and repeated-query settings where basis structure lowers per-query optimization cost or regularizes ill-conditioned parameter recovery. This suggests that their principal advantage is not universal superiority on single forward solves, but the reorganization of the inverse search space.

6. Theory, conditioning, limitations, and relation to adjacent methods

Several theoretical themes recur. The first is approximation quality. The spectral meshless formulation states that spectral approximations of analytic functions on hyperrectangles converge exponentially with truncation and reports the bound

$u(x)=C:\Phi(x),$8

for analytic $u(x)=C:\Phi(x),$9 and smooth boundaries, while also noting that Fourier extension theory provides favorable convergence on irregular subsets $\Phi(x)=\phi_1(x_1)\otimes\cdots\otimes\phi_d(x_d)$0 at the cost of orthogonality (Roggeveen et al., 29 Oct 2025). The second theme is representability or containment. In the IGA analysis, the discrete condition $\Phi(x)=\phi_1(x_1)\otimes\cdots\otimes\phi_d(x_d)$1 is central to stability and accuracy; in the neural analogue, the representability of $\Phi(x)=\phi_1(x_1)\otimes\cdots\otimes\phi_d(x_d)$2 inside the parameter space plays a similar role (Mardal et al., 2023). The third theme is identifiability: the spectral paper states explicitly that identifiability requires the chosen sampling, PDE residuals, and data constraints to eliminate parameter symmetries, and recommends regularization such as

$\Phi(x)=\phi_1(x_1)\otimes\cdots\otimes\phi_d(x_d)$3

with $\Phi(x)=\phi_1(x_1)\otimes\cdots\otimes\phi_d(x_d)$4 (Roggeveen et al., 29 Oct 2025).

Limitations are equally consistent across the literature. Exponential convergence requires smooth or analytic targets; discontinuities, sharp corners, and highly irregular boundaries degrade rates or require more modes. Nonconvex optimization can produce multiple good local minima, and first-order optimizers are less stable than Gauss–Newton or Levenberg–Marquardt in stiff inverse settings. In the offline-online method, a basis learned over a training distribution $\Phi(x)=\phi_1(x_1)\otimes\cdots\otimes\phi_d(x_d)$5 can underperform out of distribution; basis insufficiency and extrapolation outside the sampled parameter family remain explicit failure modes. In PBPK-iPINNs, success depends strongly on careful tuning of loss weights, activations, collocation, and optimizer schedules. In the peridynamic problem, naive activation choices may produce nonphysical kernels even if the loss is reduced (Roggeveen et al., 29 Oct 2025, Manor et al., 8 Sep 2025, Wickramasinghe et al., 16 Sep 2025, Difonzo et al., 2023).

The relation to adjacent methods is therefore specific rather than adversarial. Standard PINNs optimize a full network per query. FEM, SEM, and COMSOL remain superior for many pure forward problems in speed and accuracy. GPT-PINN constructs bases differently, by greedily adding basis functions that are themselves exact PINN solutions. The distinctive contribution of IP-Basis PINNs is the use of an explicit or learned basis to make inverse estimation cheaper, more stable, or more interpretable without abandoning PINN-style residual minimization (Manor et al., 8 Sep 2025, Roggeveen et al., 29 Oct 2025).

A final misconception is that basis parameterization automatically resolves inverse ill-posedness. The literature does not support that conclusion. What it does support is more limited and more technical: basis choice, containment, adaptive loss balancing, and geometry-aware collocation can materially improve conditioning, convergence, and per-query efficiency, but only when the basis matches the operator, the boundary structure, and the available data.