Partition of Unity Physics-Informed Neural Networks (POU-PINNs): An Unsupervised Framework for Physics-Informed Domain Decomposition and Mixtures of Experts
Arturo Rodriguez, Ashesh Chattopadhyay, Piyush Kumar, Luis F. Rodriguez, Vinod Kumar · 2024
Read Partition of Unity Physics-Informed Neural Networks (POU-PINNs): An Unsupervised Framework for Physics-Informed Domain Decomposition and Mixtures of Experts by Arturo Rodriguez, Ashesh Chattopadhyay, Piyush Kumar, Luis F. Rodriguez, Vinod Kumar (2024) 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-2412-06842.txt.
Introduction
Arturo Rodriguez and co-authors present a neural framework that splits physics problems by region, in work from 2024. Many physics systems change their governing rules from place to place. Standard physics-informed networks assume one rule set everywhere, which blurs those borders.
Their method divides space into subdomains, each with its own nonlinear parameters. A partition network learns the split while the physics model fits each zone. The training loss uses the physics residual itself, so no labeled data are needed. Misfit in a zone signals that its physics differ, which refines the split. Tests cover heat ablation in porous media and ice sheet flow. Accuracy gains follow from honoring local rules.
Catalog record
| CREATOR | Arturo Rodriguez, Ashesh Chattopadhyay, Piyush Kumar, Luis F. Rodriguez, Vinod Kumar |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 29 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-06842
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Partition of Unity Physics-Informed Neural Networks (POU-PINNs): An Unsupervised Framework for Physics-Informed Domain Decomposition and Mixtures of Experts Arturo Rodriguez, Ashesh Chattopadhyay, Piyush Kumar, Luis F. Rodriguez, Vinod Kumar Abstract Physics-informed neural networks (PINNs) commonly address ill-posed inverse problems by uncovering unknown physics. This study presents a novel unsupervised learning framework that identifies spatial subdomains with specific governing physics. It uses the partition of unity networks (POUs) to divide the space into subdomains, assigning unique nonlinear model parameters to each, which are integrated into the physics model. A vital feature of this method is a physics residual-based loss function that detects variations in physical properties without requiring labeled data. This approach enables the discovery of spatial decompositions and nonlinear parameters in partial differential equations (PDEs), optimizing the solution space by dividing it into subdomains and improving accuracy. Its effectiveness is demonstrated through applications in porous media thermal ablation and ice-sheet modeling, showcasing its potential for tackling real-world physics challenges. 1 Introduction Physics-informed neural networks (PINNs) are a robust framework designed to solve partial differential equations (PDEs) through optimization methods for forward problems and data-driven discovery for inverse problems [ 1 , 2 ] . While effective, PINNs face challenges in addressing parametrized PDEs with highly variable subdomains or discontinuities [ 3 ] . Standard PINNs struggle with convergence speed and accuracy because they cannot efficiently represent domain-specific features or variations [ 4 , 33 ] . To address this, we propose using partition of unity (POU) networks integrated with PINNs to form a partition of unity physics-informed neural networks (POU-PINNs). A POU network divides the computational domain into localized subdomains, ensuring features are represented where they are most relevant [ 6 , 7 ] . This decomposition is achieved by employing POU functions that sum to one and are supported over distinct spatial regions [ 8 ] . By combining the domain decomposition capabilities of partition of unity (POU) networks with the physics-informed framework of PINNs, POU-PINNs effectively address problems with significant spatial variability, such as those involving complex conductivity parameters. This hybrid approach enables precise solutions to PDEs by leveraging localized representations to capture distinct physical phenomena while narrowing the solution space [ 9 ] . 1.1 Application on Thermal Ablation In the early NASA missions, blunt nose designs were used to evenly distribute airflow and heat transfer across the spacecraft’s surface [ 10 ] . To understand this process, ablation models were created to simulate surface heating and the breakdown of heat shield materials during re-entry [ 11 ] . Based on the Darcy equation, these models included porous media flow dynamics to represent the evaporation and changes in the metal components within the heat shield’s carbon layers during thermal ablation [ 12 , 13 ] . Traditional methods for simulating thermal ablation used numerical techniques such as the finite element, finite volume, and finite difference methods aimed at solving partial differential equations (PDEs) [ 14 ] . Thermal ablation involves several coupled models, including Darcy flow, which describes porosity dynamics, and material recession, which focuses on the behavior of carbon-based materials in thermal protection systems [ 15 , 16 ] . Additionally, the Richtmyer-Meshkov instability, driven by the Rayleigh-Taylor effect, occurs along the detached bow shock wave during spacecraft re-entry [ 17 ] . To address the complexity of ablation physics, we developed a Partition of Unity Physics-Informed Neural Network (POU-PINN) solver grounded in Darcy flow principles. This solver leverages the partition of unity approach to decompose the domain into localized subdomains, enabling the accurate representation of spatially varying properties. Validated against analytical solutions from manufactured problems, the POU-PINN is implemented using TensorFlow/Keras [ 18 ] . Combining physics-based models and data efficiently reduces the solution space, enhances convergence speed, and achieves results with machine-precision accuracy. 1.2 Application on Ice-Sheet Modeling Our approach begins by using synthetically manufactured solutions designed to mimic the properties of Darcy flow, which are solved through PDE-based optimization. Manufactured solutions are simpler, allowing us to incrementally increase complexity until reaching a real-world PDE system with boundary conditions integrated into the partition of unity method presented in this paper. This approach also enables verification, as the true and exact solutions are always known within the manufactured solution framework [ 19 , 20 ] . In current studies, Darcy flow equations have been approximated using the finite element method (FEM), a state-of-the-art approach for modeling ice sheets and subsurface displacement. Sandia’s Albany code has validated FEM for over two decades, and the U.S. Department of Energy (DOE) is currently conducting experiments in Greenland and Antarctic ice shelves as a testbed for simulations performed using Albany [ 21 , 22 , 23 ] . The ultimate goal of this research is to achieve the performance level of Albany’s FEM implementation. However, current state-of-the-art sensors provide only sparse and discrete measurements of critical parameters like viscosity, which impacts conductivity and displacement values. This paper focuses on synthetic scenarios to calculate conductivity and derive scalar fields from PDE solutions as a precursor to solving real PDE problems with boundary conditions. Analytical solutions, validated using Oberkampf’s methodology, offer a baseline but fall short when compared to solving complex PDEs [ 24 , 25 ] . Although extrapolation methods can partially predict future ice sheet displacement, many parameters are dynamic and unpredictable, often resulting in chaotic bifurcations and behaviors described in studies of chaotic systems [ 26 ] . To address this unpredictability, we emphasize the importance of uncertainty quantification techniques [ 27 , 28 ] . Sensitivity analyzes, polynomial chaos expansion, reinforcement learning, Bayesian and Markov approaches, and Monte Carlo sampling are crucial for tracking and managing uncertainty as it evolves over space and time [ 29 , 30 , 31 ] . These approaches led to the discovery of qualitative physics in soft matter encounters in places where ice sheets and sub-surface flow exist, such as Greenland and the Antarctic ice shelves [ 32 , 33 , 34 ] . In this paper, we refer to conductivity to quantify the movement of ice shelves within Greenland and the Antarctic ice shelves. Over space and time, they portray the ice sheet movement in Greenland and the Antarctic, which can be used as training sets for our new method to capture qualitative physics in ice sheets. We have taken an unsupervised approach to this problem since we do not need to manually classify the different physics for the fields, such as in traditional methods. The partition of the unity network can quantify these characteristics and apply them to subdomains as partitions, then calculate the scalar Darcy fluid flow field using the POU-PINNs approach. 2 Problem Definition We have taken the forward design approach for the physics-informed neural network, where we give our partial differential equation and boundary conditions to compute the solution of our PDE and Boundary Conditions (BCs) of any rank, a vector field, or, in most of our demonstrated cases, a scalar field. First, we will describe the forward design of physics-informed neural networks. 2.1 Physics-Informed Neural Networks The generic form of a steady-state equation that is formulated as a forward problem is the following: \displaystyle\mathcal{N}[u,K] \displaystyle=0,\quad x\in\Omega (1) \displaystyle\mathcal{B}[u,K] \displaystyle=0,\quad x\in\partial\Omega (2) Where \mathcal{N} stands for