Physics-Informed Neural Networks: Unlocking Robustness, Efficiency, and Interpretability in Scientific AI
Latest 20 papers on physics-informed neural networks: Jun. 13, 2026
Physics-Informed Neural Networks (PINNs) have emerged as a powerful paradigm, blending the expressive power of deep learning with the rigorous constraints of physical laws. This synergy allows PINNs to solve complex Partial Differential Equations (PDEs), reconstruct hidden physical parameters, and model systems with sparse data, pushing the boundaries of scientific machine learning. However, challenges persist in their robustness, computational efficiency, and interpretability, particularly for complex and high-dimensional problems. Recent research, synthesized from a collection of cutting-edge papers, reveals significant breakthroughs addressing these very issues, promising to make PINNs more reliable, faster, and easier to understand.
The Big Idea(s) & Core Innovations
The core theme across recent research is the drive to make PINNs more robust and efficient in handling diverse and challenging scientific problems. A standout innovation comes from Manuel Reyna and Alexandre Tartakovsky (University of Illinois Urbana-Champaign) in their paper, “Physics-Informed Neural Networks and Radial Basis Functions for PDEs with Dirac Delta Sources”. They demonstrate a critical failure of standard PINNs when dealing with Dirac delta sources in PDEs, attributing it to the Neural Tangent Kernel (NTK) inducing global coupling. Their Radial Basis Function Residual Least Squares (RBF-RLS) method, which uses localized RBFs, successfully circumvents this issue, providing accurate solutions where PINNs fail.
Further enhancing PINN capabilities, Fateme Mohammad Mohammadi, Hector Budman, and Joshua L. Pulsipher (University of Waterloo) introduce “PL-KKT-hPINN: Enforcing Nonlinear Equality Constraints on Neural Networks via Piecewise-Linear Projection”. This method rigorously enforces nonlinear constraints by projecting neural network outputs onto feasible regions using piecewise-linear approximations. This ensures guaranteed constraint satisfaction and improved robustness, particularly in low-data regimes, a common challenge in scientific applications.
For time-dependent and high-dimensional problems, several papers offer crucial advancements. Anshima Singh and David J. Silvester (University of Manchester) extend deep adaptive sampling in “DAS-PINNs for high-dimensional partial differential equations: extending deep adaptive sampling to spacetime domains”, unifying space and time for adaptive collocation point generation. This allows PINNs to automatically focus computational effort on challenging regions, proving essential for high-dimensional PDEs where uniform sampling fails. Complementing this, Abhishek Chandra and Taniya Kapoor (KTH Royal Institute of Technology, Wageningen University & Research) introduce “Oscillatory State-Space Models as Inductive Biases for Physics-Informed Neural PDE Solvers”, an OSSM-PINN architecture that integrates oscillatory state-space dynamics as an inductive bias, achieving superior accuracy and memory efficiency for time-dependent PDEs by better capturing modal structures.
When attention meets accuracy, Brandon Yee et al. (Physics Lab, Yee Collins Research Group) present the “Multi-Scale Attention Transformer for PDE Solving on Irregular Domains”, named MSAT. This architecture, specifically for irregular domains, leverages parallel attention streams to encode spatiotemporal solutions, outperforming Fourier-based neural operators and providing principled guidance on when to use attention over spectral methods. Similarly, Nanxi Chen et al. (Tongji University, University of Oxford, Yale University) investigate “On the training of physics-informed neural operators for solving parametric partial differential equations”, finding that PI-CViT (a CViT-based PINO) consistently outperforms other neural operator backbones and that well-resolved physics-informed training can match or exceed purely data-driven approaches.
Addressing the ill-posed nature of inverse problems, Mahmoud Elhadidy et al. (University of Utah) in “Wall Shear Stress Reconstruction from Concentration: Differentiable Physics and Physics-Informed Neural Networks” demonstrate that differentiable physics (hard constraints) often outperforms PINNs (soft constraints) for inverse hemodynamic problems like wall shear stress reconstruction, especially with far-field measurements. This highlights the importance of rigorous constraint enforcement. However, for a different class of inverse problems in materials science, Pouria Behnoudfar et al. (University of Wisconsin-Madison) introduce “SPLIT-PINN: Separable Probability Learning Technique via Physics-Informed Neural Networks for High-Dimensional Probabilistic Modeling”, which effectively infers high-dimensional drift fields in Liouville equations from PDF data by using a marginal-correction drift decomposition with orthogonality constraints.
Improving training efficiency and stability are also critical. Ryo Sagawa et al. (The University of Osaka) accelerate SAV-based optimization with “Accelerating SAV-based optimization via randomized low-rank Hessian approximation”, achieving significant speedups for PINN training by incorporating approximate Hessian information. Yingjie Shao et al. (Wageningen University & Research) tackle numerical precision with a “Curvature-aware dynamic precision approach for physics-informed neural networks” that adaptively switches between FP32 and FP64 based on L-BFGS curvature, balancing accuracy and computational cost. Further, Zhengqi Zhang and Jing Li (Ant Group, Zhejiang Lab) optimize fractional PINNs in “Transformed Diffusion-Wave fPINNs” by transforming the Caputo-residual to reduce the automatic-differentiation burden, yielding substantial memory and FLOP savings. For non-homogeneous boundary conditions, Dongjie Liu et al. (Beijing University of Technology, Chongqing University of Technology) introduce “Tr-PINNs for 2D Incompressible Navier-Stokes Equations”, using trace theorem and H^{1/2}-norm for boundary residuals to achieve remarkable accuracy improvements.
Finally, addressing parametric studies, Anna Lazareva and Alexander Tarakanov (HSE University, VK and HSE University) propose “Loss-Conditional PINNs for Parametric PDE Families”, LC-PINN, which learns a continuous family of solutions by treating loss weights and PDE parameters as conditioning variables, amortizing computation across parametric PDE families. The crucial aspect of interpretability is addressed by Aleksander Krasowski et al. (Fraunhofer HHI, Technische Universität Berlin, BIFOLD, Technological University Dublin) with “PINNfluence: Interpreting PINNs through Influence Functions”, a framework that adapts influence functions to PINNs to attribute predictions to individual training points and loss components, providing vital diagnostics for understanding model behavior and quality.
Under the Hood: Models, Datasets, & Benchmarks
These advancements are underpinned by novel architectural designs, clever use of existing tools, and rigorous benchmarking:
- RBF-RLS Method: Utilizes Radial Basis Functions (RBFs) for localized influence, contrasting with global coupling issues in standard PINNs for Dirac delta sources. Demonstrated on transport problems in porous media and rivers.
- PL-KKT-hPINN: Extends KKT-hPINN with piecewise-linear projection for enforcing nonlinear equality constraints. Validated on 1D and 2D Continuous Stirred-Tank Reactor (CSTR) case studies. Code available at https://github.com/PULSI-Opt/PL-KKT-hPINN.
- Adaptive KANs (RGA KANs): Spyros Rigas et al. (National and Kapodistrian University of Athens, Academy of Athens, Greece) in “An adaptive framework for the axisymmetric pulsar magnetosphere using physics-informed Kolmogorov-Arnold networks” replaced standard MLPs with Residual-Gated Adaptive Kolmogorov-Arnold Networks (RGA KANs) for solving the axisymmetric pulsar magnetosphere problem. This, combined with an automated adaptive training pipeline (Learning Rate Annealing and Residual-Based Attention), achieved two orders of magnitude improvement in MSE. Their framework, PulsarX, is open-sourced in JAX (https://github.com/).
- DAS-PINNs with KRnet: Extends Deep Adaptive Sampling to spacetime domains using a Normalising Flow Neural Network (KRnet) to approximate residual distributions for adaptive collocation point generation. Evaluated on benchmark problems from 2D to 8D spatial dimensions.
- OSSM-PINN: Incorporates oscillatory state-space dynamics (LinOSS cell) with a PDE-aware spectral basis. Benchmarked on forward, inverse, high-dimensional (up to 100D Schrödinger), non-rectangular, and large-domain problems.
- PI-MFA: Junoh Jung et al. (Argonne National Laboratory) introduce “A Physics-Informed B-Spline Framework for Continuous Approximation of Flow Data” which reconstructs discrete flow data using tensor-product B-spline representations with embedded PDE, BC, and IC constraints, outperforming PINNs on convection-diffusion, Burgers, and Navier-Stokes equations.
- MSAT (Multi-Scale Attention Transformer): A novel attention-based architecture for irregular domains. Evaluated against 9 models across 5 PINNacle benchmarks (https://arxiv.org/abs/2306.08827).
- PI-CViT: A CViT (Cross-attention Vision Transformer)-based neural operator, found to be the strongest performing PINO backbone for parametric PDEs (Helmholtz, Schrödinger, Burgers, Buckley-Leverett equations). Code released with paper at https://github.com/NanxiiChen/PI-CViT.
- Differentiable Physics (DiffFEM): Yongjin Choi, Hyeonbin Moon, and Seunghwa Ryu (KAIST, Georgia Institute of Technology) critically evaluate PINNs against Differentiable Finite Element Method (DiffFEM) for FWD backcalculation in “Critical evaluation of PINN for FWD inverse analysis and differentiable FEM as an alternative”. DiffFEM, using hard physics constraints via FEM and automatic differentiation, proves more robust and accurate for systems with sharp material discontinuities, showcasing a ~600x speedup over XPINN.
- SPLIT-PINN: Employs marginal-correction drift decomposition with orthogonality constraints and Residual-Based Adaptive Distribution (RAD). Applied to real datasets of polycrystal state evolution (von Mises stress, dislocation density, equivalent plastic strain rate).
- Tr-PINNs: Utilizes trace theorem with H^{1/2}-norm for boundary residuals in 2D incompressible Navier-Stokes equations. Validated with Taylor-Green vortex test cases, comparing against OpenFOAM CFD.
- MAdam (Metric-Aware Multi-Objective Adam): A drop-in wrapper that preconditions multi-objective Adam with a preference-conditioned diagonal Fisher Information Matrix. Evaluated across multi-task learning, Pareto-front recovery, and PINN benchmarks (PINNacle suite, https://arxiv.org/abs/2406.12167).
- Curvature-aware Dynamic Precision: Reuses L-BFGS curvature information for adaptive FP32/FP64 switching. Validated across MLP, PINNsFormer, PINNMamba, and KAN architectures.
- tDWfPINNs: Transforms Caputo-residual for time-fractional diffusion-wave equations, evaluated with Monte Carlo and Gauss-Jacobi quadratures. Focuses on 1D and 2D fractional PDEs.
- LC-PINN: A loss-conditional PINN architecture using both concatenation and FiLM conditioning. Empirically evaluated across Helmholtz, Schrödinger, Burgers, and Buckley-Leverett equations.
- PINNFLUENCE: Interprets PINNs using adapted influence functions. Diagnoses well-trained vs. poorly-trained models across five time-dependent PDEs. Code available at https://github.com/aleks-krasowski/pinnfluence.
Impact & The Road Ahead
The collective impact of this research is profound. These advancements are pushing PINNs from promising theoretical tools to practical, robust, and efficient solvers for real-world scientific and engineering challenges. The ability to handle Dirac delta sources, enforce hard nonlinear constraints, adaptively sample in high dimensions, and leverage oscillatory inductive biases means PINNs can now tackle a broader spectrum of complex physical phenomena with greater accuracy and stability.
Improved training methodologies, such as curvature-aware dynamic precision and accelerated optimization, make PINNs more computationally viable for large-scale applications. The advent of neural operators that can learn entire families of solutions (LC-PINN) or generalize across irregular geometries (MSAT, PI-CViT) signals a shift towards more versatile and powerful surrogate models.
The development of “PINNfluence” is particularly exciting, as it addresses a critical gap in the interpretability of PINNs. Understanding why a PINN makes certain predictions and how different physics components influence the solution is vital for building trust and enabling debugging in high-stakes scientific applications like medical imaging, climate modeling, and material science.
Looking ahead, we can anticipate further integration of these techniques. The push for hard constraints, as seen in DiffFEM and PL-KKT-hPINN, will likely continue, especially for problems requiring high fidelity and guaranteed physical consistency. The exploration of new neural architectures like KANs and attention-based transformers will yield even more expressive and efficient models. The continuous development of adaptive sampling and optimization strategies will be crucial for scaling PINNs to ever-larger and more complex problems. Ultimately, these breakthroughs are paving the way for a new era of scientific discovery, where AI and physics collaborate seamlessly to unlock insights previously unattainable. The future of physics-informed AI is brighter and more robust than ever before.
Share this content:
Post Comment