Physics-Informed Neural Networks: Pushing the Boundaries of Scientific Machine Learning
Latest 9 papers on physics-informed neural networks: Jul. 11, 2026
Physics-InInformed Neural Networks (PINNs) have emerged as a powerful paradigm, blending the expressive power of neural networks with the robust principles of physics. By embedding governing equations, boundary conditions, and domain characteristics directly into the neural network’s loss function, PINNs offer a mesh-free, data-efficient approach to solving complex scientific and engineering problems. This ability to infuse deep learning with physical laws is driving breakthroughs across various disciplines, tackling challenges from fluid dynamics to inverse problems. Let’s dive into some of the latest advancements that are significantly pushing the boundaries of what PINNs can achieve.
The Big Idea(s) & Core Innovations
The fundamental challenge in many scientific machine learning tasks, especially with PINNs, lies in achieving both accuracy and stability, particularly in complex or ill-conditioned scenarios. Recent research is addressing this head-on with novel optimization strategies, specialized architectures, and robust solution methodologies.
One major theme revolves around enhancing the optimization landscape for PINNs. Researchers from the Mathematical Institute, University of Oxford, in their paper, “An Optimisation Framework for the Well-Conditioned Training of Physics-Informed Neural Networks”, introduce DSGNAR (Doubly-Sketched Gauss–Newton with Adaptive Ratio). This ground-breaking framework directly tackles ill-conditioning by employing a doubly-sketched Gauss-Newton approach with adaptive regularization, achieving unprecedented accuracy (errors as low as 3×10⁻¹⁶) by prioritizing conditioning. Complementing this, Jianing Liu and Dong H. Zhang from the University of Science and Technology of China propose “Higher-Order Geometric Updates for Levenberg-Marquardt Method via Riemann Normal Coordinates”. Their RNC-LM method extends geodesic acceleration to arbitrary orders, providing more geometrically consistent updates that avoid PINN failure modes and significantly speed up tasks like potential-energy-surface fitting. Furthermore, a team from the University of Münster and The Chinese University of Hong Kong, in “Learning rate adaptive stochastic gradient descent optimization methods: numerical simulations for deep learning methods for partial differential equations and convergence analyses”, addresses the notoriously tricky learning rate tuning with an adaptive SGD approach. Their method adjusts learning rates dynamically based on empirical loss estimates, demonstrating faster convergence for PINNs and other deep learning PDE solvers without manual tuning, and providing rigorous convergence proofs.
Another crucial innovation is the development of specialized architectures and methods for complex problem types. For instance, Qian Hua et al. from Fujian University of Technology tackle the persistent issue of negative transfer in PINN inverse problems with “Target-Guided Selective Reweighting for Physics-Informed Neural Network Inverse Problems: A Transfer Learning Approach”. Their TGSR-PINN uses neuron-level representation correction to ensure accurate physical parameter recovery, which direct fine-tuning often misses. For problems with sharp interfaces and discontinuities, Jingnan Yao et al. from Harbin Institute of Technology and Michigan State University introduce “Level-set physics-informed neural networks for domain inverse problems of gravimetry”. This framework combines level-set representations with an interface-aware backpropagation strategy to overcome gradient issues near discontinuities, successfully recovering subsurface density distributions. Similarly, to master convection-dominated problems that create notoriously thin layers, Zihao Guo, Xin Li, and Zhihong Xia from Great Bay University and Northwestern University developed “LRX-PINN: A Layer-Resolving XNet Physics-Informed Neural Network with Integrated Cauchy Activations for Convection-Dominated Problems”. Their LRX-PINN uses integrated Cauchy activations that inherently match the value-derivative structure of thin layers, achieving superior accuracy with significantly fewer parameters. Beyond Euclidean spaces, Yufang Jiang et al. from Nanjing Normal University and Nanjing University extend PINN applicability to high-dimensional Riemannian manifolds with “Domain decomposition methods with Physics-informed neural networks for elliptic equations on manifolds”. By combining PINNs with domain decomposition, they mitigate the curse of dimensionality for complex geometries, enabling solutions for elliptic equations on manifolds up to 10 dimensions without meshing. Finally, for multiphase flow simulations in porous media, Yuanshuo Kong et al. from Shandong University introduce “The PICNN-Assisted Physics-Preserving Scheme for Thermodynamically Consistent Two-Phase Flow in Porous Media”. This method uses a CNN for prediction, followed by a physics-preserving correction step, ensuring energy stability, mass conservation, and saturation bounds at machine precision, even with discontinuous permeability fields.
Under the Hood: Models, Datasets, & Benchmarks
These advancements are often powered by innovative architectural designs, strategic data utilization, and robust benchmarking, enabling PINNs to tackle previously intractable problems:
- DSGNAR (Doubly-Sketched Gauss–Newton with Adaptive Ratio): This optimization framework, detailed in “An Optimisation Framework for the Well-Conditioned Training of Physics-Informed Neural Networks”, leverages CountSketch for row sketching and Subsampled Randomized Cosine Transform (SRCT) for column sketching to efficiently compute Gauss-Newton steps. It’s validated across a spectrum of PDEs, including Burgers, Kuramoto-Sivashinsky, and Navier-Stokes equations, showcasing remarkable accuracy improvements.
- RNC-LM: Featured in “Higher-Order Geometric Updates for Levenberg-Marquardt Method via Riemann Normal Coordinates”, this method modifies the Levenberg-Marquardt optimizer by integrating Riemann normal coordinates for higher-order geodesic updates. It’s benchmarked on large-scale potential-energy-surface fitting (achieving 34x speedup) and challenging reaction-diffusion PINN problems.
- TGSR-PINN: This transfer learning approach for inverse problems, from “Target-Guided Selective Reweighting for Physics-Informed Neural Network Inverse Problems: A Transfer Learning Approach”, focuses on neuron-level scoring using Taylor sensitivity and pre-activation variance, combined with Gaussian Mixture Models (GMM). Its effectiveness is demonstrated on high-Péclet advection-diffusion, Allen-Cahn to Burgers cross-PDE-family transfer, and noisy reaction-diffusion tasks. Code available: https://github.com/HuaQian-TGSR/TGSR-PINN
- Level-set PINNs: Introduced in “Level-set physics-informed neural networks for domain inverse problems of gravimetry”, this framework uses level-set representations to handle discontinuous interfaces and an interface-aware backpropagation strategy. It’s validated with extensive 2D and 3D numerical experiments for gravimetry inverse problems. Software available: https://doi.org/10.5281/zenodo.20790726
- LRX-PINN: As discussed in “LRX-PINN: A Layer-Resolving XNet Physics-Informed Neural Network with Integrated Cauchy Activations for Convection-Dominated Problems”, this architecture integrates Cauchy activation functions into an XNet-like framework to resolve thin layers in convection-dominated problems. It outperforms PIKAN and Fourier-feature PINNs on various benchmarks, including those with interior, boundary, and curved layers. Code available for baselines: https://github.com/wanjiashan/PIKANs, https://github.com/songc0a/Fourier-feature-PINN-based-multifrequency-multisource-Helmholtz-s, https://github.com/airexworklab/fastvpinns
- PINNs for Elastodynamic Wave Propagation: The framework presented by Sonal Ankush Chibire et al. from Northern Illinois University in “A Physics-Informed Neural Network Framework for Elastodynamic Wave Propagation in Bimaterial Systems” uses PINNs to solve axisymmetric linear elasticity equations. It’s trained and validated against high-fidelity ANSYS Explicit Dynamics simulations for steel-aluminum bimaterial systems, serving as a continuous surrogate model for Split Hopkinson Pressure Bar analysis.
- Domain Decomposition PINNs: For elliptic equations on manifolds, “Domain decomposition methods with Physics-informed neural networks for elliptic equations on manifolds” employs traditional Domain Decomposition Methods (DDMs) with PINNs. Numerical validation includes high-dimensional spheres, product manifolds, and complex projective spaces (CP3).
- PICNN-Assisted Scheme: The work in “The PICNN-Assisted Physics-Preserving Scheme for Thermodynamically Consistent Two-Phase Flow in Porous Media” uses a Convolutional Neural Network (CNN) for initial prediction, followed by a mGSAV-LM (modified Gradient-Scaling and Volume Averaging Levenberg-Marquardt) framework for physics-preserving correction. It is validated through manufactured-solution tests and heterogeneous porous media simulations.
- Adaptive Learning Rate SGD: The adaptive Adam optimizer from “Learning rate adaptive stochastic gradient descent optimization methods: numerical simulations for deep learning methods for partial differential equations and convergence analyses” demonstrates faster convergence for Deep Kolmogorov methods, PINNs (e.g., sine-Gordon, Allen-Cahn PDEs), and Deep Ritz methods. Code available: https://github.com/deeplearningmethods/adaptive-learning-rate
Impact & The Road Ahead
These advancements significantly broaden the scope and reliability of Physics-Informed Neural Networks. The ability to achieve machine-precision accuracy in complex PDE solutions (“An Optimisation Framework for the Well-Conditioned Training of Physics-Informed Neural Networks”), to robustly recover physical parameters in inverse problems (“Target-Guided Selective Reweighting for Physics-Informed Neural Network Inverse Problems: A Transfer Learning Approach”), and to tackle extreme challenges like convection-dominated flows (“LRX-PINN: A Layer-Resolving XNet Physics-Informed Neural Network with Integrated Cauchy Activations for Convection-Dominated Problems”) or multi-material wave propagation (“A Physics-Informed Neural Network Framework for Elastodynamic Wave Propagation in Bimaterial Systems”) signals a maturing field. The integration of domain decomposition for high-dimensional manifolds (“Domain decomposition methods with Physics-informed neural networks for elliptic equations on manifolds”) and robust handling of discontinuities through level-set methods (“Level-set physics-informed neural networks for domain inverse problems of gravimetry”) opens doors for new applications in geophysics, materials science, and beyond.
The future of PINNs lies in even more sophisticated physics-neural integration, moving towards adaptive methodologies that dynamically respond to problem complexity. The drive for efficient and stable optimization, as seen with adaptive learning rates (“Learning rate adaptive stochastic gradient descent optimization methods: numerical simulations for deep learning methods for partial differential equations and convergence analyses”) and higher-order geometric updates (“Higher-Order Geometric Updates for Levenberg-Marquardt Method via Riemann Normal Coordinates”), will continue. We can anticipate hybrid approaches, combining PINNs with traditional numerical methods to leverage the strengths of both, ultimately paving the way for truly intelligent and trustworthy scientific discovery. The progress is clear: PINNs are no longer just an academic curiosity; they are becoming an indispensable tool in the scientific machine learning toolkit, poised to revolutionize how we understand and simulate the physical world.
Share this content:
Discover more from SciPapermill
Subscribe to get the latest posts sent to your email.
Post Comment