Physics-informed KAN PointNet: Deep learning for simultaneous solutions to inverse problems in incompressible flow on numerous irregular geometries
Ali Kashefi, Tapan Mukerji · 2025
Read Physics-informed KAN PointNet: Deep learning for simultaneous solutions to inverse problems in incompressible flow on numerous irregular geometries by Ali Kashefi, Tapan Mukerji (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-06327.txt.
Introduction
Ali Kashefi and Tapan Mukerji build a physics taught point net that solves back flow asks on many odd shapes at once, in a 2025 paper. New arc nets beat plain nets in some physics tasks of late. Past arc nets solve one shape per teach run, a curb on their use for shape pools.
Shared arc nets sit in a point net to read shape traits in the build. Loss terms mix law scraps, sparse sight marks, and part wall facts. Bend curve kinds trade cost and truth in the shared build. A heat flow case with a shape shift core tests the frame. Truth beats the plain point net at like cost.
Catalog record
| CREATOR | Ali Kashefi, Tapan Mukerji |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 72 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-06327
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Physics-informed KAN PointNet: Deep learning for simultaneous solutions to inverse problems in incompressible flow on numerous irregular geometries Ali Kashefi, Tapan Mukerji Abstract Kolmogorov-Arnold Networks (KANs) have gained attention as a promising alternative to traditional Multilayer Perceptrons (MLPs) for deep learning applications in computational physics, especially within the framework of physics-informed neural networks (PINNs). Physics-informed Kolmogorov-Arnold Networks (PIKANs) and their variants have been introduced and evaluated to solve inverse problems. However, similar to PINNs, current versions of PIKANs are limited to obtaining solutions for a single computational domain per training run; consequently, a new geometry requires retraining the model from scratch. Physics-informed PointNet (PIPN) was introduced to address this limitation for PINNs. In this work, we introduce physics-informed Kolmogorov-Arnold PointNet (PI-KAN-PointNet) to extend this capability to PIKANs. PI-KAN-PointNet enables the simultaneous solution of an inverse problem over multiple irregular geometries within a single training run, reducing computational costs. We construct KANs using Jacobi polynomials and investigate their performance by considering Jacobi polynomials of different degrees, as well as special cases such as Legendre polynomials, Chebyshev polynomials of the first and second kinds, and Gegenbauer polynomials, in terms of both computational cost and prediction accuracy. As a benchmark test case, we consider natural convection in a square enclosure with a cylinder, where the cylinder’s shape varies across a dataset of 135 geometries. We compare the performance of PI-KAN-PointNet with that of PIPN (i.e., physics-informed PointNet with MLPs) and observe that, with approximately an equal number of trainable parameters and similar computational cost, PI-KAN-PointNet provides more accurate predictions. Finally, we explore the combination of KAN and MLP in constructing a physics-informed PointNet. Our findings indicate that a physics-informed PointNet model employing MLP layers as the encoder and KAN layers as the decoder represents the optimal configuration among all models investigated. [orcid=0000-0003-0014-9051] [orcid=0000-0003-1711-1850] keywords ,Kolmogorov-Arnold networks ,PointNet ,Physics-informed deep learning ,Inverse problems ,Irregular geometries ,Steady state incompressible flow † † corresponding: Corresponding author 1 Introduction and motivation Physics-informed Neural Networks (PINNs), first introduced by Raissi et al. in 2019 [ 1 ] , are increasingly recognized as a valuable tool for solving inverse problems in diverse scientific and industrial domains. These include solid mechanics [ 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 ] , both incompressible and compressible flow [ 19 , 20 , 21 , 22 , 23 , 1 , 24 , 25 , 26 , 27 ] , chemistry [ 28 , 29 ] , heat transfer [ 20 , 30 , 31 ] , and flow in porous media [ 32 , 33 , 34 ] . The central idea behind PINNs for inverse problems is to train a neural network to minimize both the residuals of the governing partial differential equations and the distance between the solution provided by the neural network and sparse observational data at sensor locations in specific norms such as the Euclidean norm. Additional constraints, such as boundary or initial conditions, may also be incorporated into this optimization process. The choice of the neural network architecture in PINNs significantly influences the performance and capability of PINN models. A prevalent choice is the fully connected neural network [ 3 , 2 , 35 , 36 , 37 , 38 ] . However, a notable limitation of using fully connected networks is that they are generally suited to solving inverse problems for a single specific geometry. For each new geometry, the network must be retrained from the ground up, which incurs significant computational costs. This challenge is particularly pronounced when exploring a broad range of geometric parameters for optimizing industrial designs. The issue of high computational expense in such scenarios was addressed by [ 39 , 20 , 40 ] . To resolve this issue, Kashefi and Mukerji 2022 introduced Physics-informed PointNet (PIPN). In PIPN, instead of using a fully connected neural network, PointNet [ 41 ] is employed. Since PointNet [ 41 ] and its advanced versions [ 42 , 43 ] can capture the geometric features of inputs in a latent space, they can differentiate between various geometries. As a result, inverse problems across multiple geometries can be solved simultaneously in a single training process. Kashefi and Mukerji 2022 solved an inverse heat transfer convection problem over 108 irregular geometries simultaneously. Similarly, Kashefi et al. 2023 used PIPN to solve an inverse linear elasticity problem over 532 domains with irregular geometries. The architecture of PointNet [ 41 ] , derived from the field of computer graphics, is more complex than fully connected neural networks and involves several components. The primary components of PointNet are fundamentally built using shared Multilayer Perceptrons (MLPs) [ 44 ] . In this article, we introduce Physics-Informed Kolmogorov-Arnold PointNet (PI-KAN-PointNet), where the network replaces MLPs with Kolmogorov-Arnold Networks (KANs) [ 45 , 46 ] . KANs [ 45 , 46 ] have recently been introduced as a novel alternative to conventional MLPs [ 47 , 48 , 49 ] . KANs are rooted in the Kolmogorov-Arnold representation theorem [ 50 , 51 , 52 , 53 , 54 , 55 , 56 , 57 ] . Unlike MLPs, which rely on training weights and biases with fixed activation functions, KANs focus on training the activation functions themselves [ 45 ] . KANs have been incorporated into Convolutional Neural Networks (CNNs) [ 58 , 59 ] , graph neural networks [ 60 , 61 , 62 , 63 ] , and PointNet [ 64 , 65 ] . This approach has proven effective across a variety of fields, including physics-informed machine learning [ 66 , 67 , 68 , 69 , 70 , 71 , 72 , 73 ] , deep operator networks [ 74 , 67 ] , neural ordinary differential equations [ 75 , 76 ] , image classification [ 58 , 77 , 78 , 79 , 80 , 81 , 82 , 83 , 84 , 85 ] , image segmentation [ 86 , 87 ] , image detection [ 88 ] , audio classification [ 83 ] , and numerous other scientific and industrial applications [ 89 , 90 , 91 , 92 , 93 , 94 , 95 , 96 , 97 , 98 , 99 , 100 , 101 , 102 , 103 , 104 , 105 , 106 , 107 , 108 ] . In this article, we focus on the contributions of KANs to the field of physics-informed deep learning. Several researchers have made significant advancements in the application of KANs in physics-informed deep learning, particularly in solving inverse problems. Wang et al. 2024 introduced a physics-informed Kolmogorov-Arnold network for solving inverse problems in solid mechanics, such as nonlinear hyperelasticity. Shukla et al. 2024 used a physics-informed Kolmogorov-Arnold network with Chebyshev polynomials as the basis for KANs to simulate two-dimensional lid-driven cavity problems and solve the Allen-Cahn equation. Similarly, Toscano et al. 2024 applied the same version of the physics-informed Kolmogorov-Arnold network to infer velocity and temperature fields in three-dimensional turbulent flow. Howard et al. 2024 developed a domain decomposition algorithm to reduce the computational cost