Physics-Informed Neural Networks: Scaling, Solving Complex Physics, and Getting it Right
Latest 16 papers on physics-informed neural networks: Oct. 10, 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 to solve complex scientific and engineering problems. Yet, as the field matures, researchers are pushing beyond foundational concepts, tackling critical challenges like scalability, accuracy for complex phenomena, and robust uncertainty quantification. Recent breakthroughs, highlighted in a collection of new papers, reveal innovative solutions to these hurdles, promising to unlock PINNs’ full potential.
The Big Idea(s) & Core Innovations
The central theme across these advancements is overcoming limitations of standard PINNs to enhance their accuracy, robustness, and applicability to real-world, often complex, scenarios.
A significant hurdle for PINNs is accurately capturing sharp discontinuities or fine structures. Researchers from the Max Planck Institute for Plasma Physics and Université de Strasbourg, in their paper “A neural characteristic mapping method: Lagrangian PINNs based on flow maps for transport-dominated problems”, introduce a novel Lagrangian PINN that approximates flow maps instead of the solution itself. This brilliant conceptual shift allows the complexity of fine structures to be transferred from the network’s direct representation to the composition of simpler ‘submaps’ that remain close to identity, significantly improving resolution for transport-dominated problems like incompressible Euler equations.
Another major challenge lies in the often-overlooked subtleties of uncertainty quantification. Michael Obermayr and Robert Peharz from TU Graz, in “Uncovering and Fixing Collider Bias in Bayesian PINNs”, reveal a critical systematic posterior bias in standard Bayesian PINNs (B-PINNs). This “collider bias” leads to overconfident and incorrect parameter estimates. Their hierarchical chain model, combined with Particle MCMC, eliminates this bias by properly accounting for physics-dependent normalization constants, a vital step toward reliable uncertainty estimates in scientific ML.
For complex multi-physics systems, ensuring global conservation is paramount. “Cova-PINN: Cross-Domain Conservation Physics-Informed Neural Network for Fluid-Solid Conjugate Heat Transfer in Complex Geometries” by Weizheng Zhang et al. from Shandong University addresses the “conservation gap” in fluid-solid conjugate heat transfer. They show that pointwise training can yield plausible local temperature fields but fail at accurate global energy transfer. Cova-PINN aligns conservation support with the complete thermal interaction path through cross-domain control volumes and paired-wall closure, leading to substantial improvements in device-level energy conservation for heat exchangers.
Addressing the notorious spectral bias of PINNs—their difficulty in capturing high-frequency components like shocks—Muhammad M. Akmal et al. from The University of Texas at Austin introduce “Gen-PINNs: Generative Adversarial Physics Informed Neural Networks for solving partial differential equations”. This framework leverages generative adversarial learning, where a generator learns PDE solutions and discriminators evaluate physics-based residual features, alongside Fourier input embeddings and dynamic loss weighting, to significantly improve accuracy on stiff and shock-front PDEs. Complementing this, Akshay Thakur and Matthew J. Zahra from the University of Notre Dame propose a “Feature tracking in physics-informed neural networks via joint optimization of nonlinear deformation manifolds: application to shocks”. This FT-PINN framework uses a jointly optimized deformation map to automatically concentrate collocation points along shocks, avoiding the need for explicit resampling and resolving features of arbitrary geometry.
Scalability and optimization are also key. The paper “PE-EK-PINN: Physics Embedding with Evolving Kernel for Scalable Physics-Informed Neural Networks” by Huiwen Zhang et al. from the University of Wisconsin-Madison introduces a hierarchical framework that reuses learned subsystem fields as “physics kernels” to dramatically reduce training costs (from O(N) to O(log N)) for large-scale wave field modeling. On the optimization front, Joohwan Ko et al. from the University of Massachusetts Amherst present “SoftServe: A Scalable Quasi-Newton Method for Deep Learning”. SOFTSERVE is a quasi-Newton framework that provides positive-definite curvature estimates without line searches, using GPU-friendly Newton-Schulz iterations and Kronecker-factored approximations, demonstrating superior performance on ill-conditioned tasks including PINNs.
For high-dimensional parametric PDEs, Denis Korolev and Martin Eigel from WIAS Berlin propose “Tensor-Train Compressed Separable PINNs: A Curvature-Aware Optimization Framework for Parametric PDEs in High Dimensions”. By exploiting coordinate-separable neural architectures and differential operators, this framework achieves exact compressed formulations of the Gauss-Newton step, allowing it to tackle the curse of dimensionality in scenarios like uncertainty quantification.
Finally, ensuring robust training for specific applications is crucial. “Hyperparameter selection for equation learning with biologically-informed neural networks” by William Lavery et al. from Uppsala University provides a ground-truth-free diagnostic workflow for Biologically-Informed Neural Networks (BINNs), a PINN subclass, for learning PDEs from sparse, noisy biological data. Their rules of thumb for network capacities and early-stopping patience make BINN training systematic. Furthermore, in “Operator-informed initialization for Fourier features physics-informed neural networks”, Juan Molina et al. from the National Center for Artificial Intelligence, Chile, address operator-induced spectral bias by proposing a novel PDE-informed initialization strategy that tailors weight distributions to the PDE’s differential operator, balancing convergence rates across the frequency spectrum.
Under the Hood: Models, Datasets, & Benchmarks
These papers not only introduce innovative methodologies but also advance the tools and practices for robust PINN development:
- Lagrangian PINNs based on Flow Maps: This method is implemented within the
scimbalibrary, providing a software framework for their novel submap decomposition and sequential training approach. (A neural characteristic mapping method: Lagrangian PINNs based on flow maps for transport-dominated problems) - Collider Bias in B-PINNs: The authors provide code at https://github.com/obermayr/ColliderBias for their hierarchical chain model and Particle Marginal Metropolis-Hastings (PMMH) inference, validated on damped oscillators, double pendulums, and relativistic particle systems.
- Cova-PINN: Benchmarked on complex TPMS (Primitive, Gyroid, Diamond, IWP) and DualMS heat exchanger geometries, with reference solutions from ANSYS CFX 2023 R1. (Cova-PINN: Cross-Domain Conservation Physics-Informed Neural Network for Fluid-Solid Conjugate Heat Transfer in Complex Geometries)
- Gen-PINNs: Evaluated against challenging PDEs like Burgers, Allen-Cahn, and Kuramoto-Sivashinsky equations, known for their sharp-front and shock behaviors. (Gen-PINNs: Generative Adversarial Physics Informed Neural Networks for solving partial differential equations)
- M-NEM (Mesh-Based Neural Energy Method): This extension of the neural energy method (NEM) uses mesh-based discretization and RBFNNs (Radial Basis Function Neural Networks) with Matérn-3/2 kernels, outperforming MLPs and KANs. It leverages FEniCSx, Basix, and UFL for finite element integration, with PyTorch code to be released. (A mesh-based neural energy method for the simulation of heterogeneous composites)
- KAN-PINNs for Biharmonic Equation: Utilizes Kolmogorov-Arnold Neural Networks (KANs) with learnable univariate functions on network edges, optimized with a hybrid AdamW-to-L-BFGS pipeline, to solve fourth-order elliptic PDEs. (Mesh-Free Numerical Approximation of the Biharmonic Equation via Optimized Kolmogorov–Arnold Neural Networks)
- Domain-informed Adaptive Sampling: Employs Conditional Flow Matching to learn process condition-dependent high-residual distributions in metal additive manufacturing thermal modeling, validated using high-fidelity FEM simulations from JAX-AM. Code is available at https://github.com/hsleecri/AdaptiveSamplingPINN-AM. (Domain-informed Adaptive Sampling for Generalizable PINNs in Metal Additive Manufacturing via Conditional Flow Matching)
- L0-Neural DDE (L0-NDDE): This method for identifying drivers and dynamics of delayed physical systems uses L0 regularization. It is benchmarked on the CausalDynamics dataset and the Dysts benchmark for chaotic systems. Code is provided at https://anonymous.4open.science/r/ODEDriverIdentifiability-C325/README.md. (Identifiability Guarantees for Drivers and Dynamics of Delayed Physical Systems)
- CI-PINN (Causal Integral PINN): Employs a Causal Integral Neural Network (CinNet) architecture with Volterra-type causal integral terms, demonstrating superior accuracy and robustness on evolution equations like Allen-Cahn, KdV, and Cahn-Hilliard. (CI-PINN: Causal Integral Physics-Informed Neural Network for Solving Evolution Equations)
Impact & The Road Ahead
These recent advancements significantly broaden the scope and reliability of PINNs. The ability to accurately model transport-dominated problems, reliably quantify uncertainty, ensure conservation in multi-physics systems, and capture sharp discontinuities means PINNs can now tackle a wider array of complex, real-world scientific and engineering challenges with greater confidence. From designing efficient heat exchangers and simulating composite materials to understanding complex biological systems and optimizing additive manufacturing processes, the practical implications are vast.
The theoretical work on error bounds for Hamilton-Jacobi-Bellman equations (Convergence of kernel and neural-network methods for Hamilton–Jacobi–Bellman equations on unbounded domains by Yumiharu Nakano from Institute of Science Tokyo) provides a crucial foundation for rigorous application and further development of PINNs in optimal control. Similarly, the insights into operator-informed initialization and robust hyperparameter selection pave the way for more stable and efficient training of these models. The development of scalable optimization frameworks like SoftServe and hierarchical kernel learning like PE-EK-PINN addresses the crucial need for PINNs to handle increasingly large and complex systems.
The road ahead for PINNs looks incredibly promising. Future research will likely focus on further integrating these innovations into unified, general-purpose frameworks. We can anticipate more sophisticated architectural designs that inherently encode physics, better methods for handling diverse data types and noise, and continued advancements in optimization and uncertainty quantification. As these tools become more robust and accessible, PINNs are poised to become an indispensable component of scientific discovery and engineering innovation, truly bridging the gap between data-driven AI and first-principles physics.
Share this content:
Discover more from SciPapermill
Subscribe to get the latest posts sent to your email.
Post Comment