Digitala Vetenskapliga Arkivet

Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
MILP initialization for solving parabolic PDEs with PINNs
KTH, School of Electrical Engineering and Computer Science (EECS), Intelligent systems, Decision and Control Systems (Automatic Control).ORCID iD: 0009-0005-1869-3495
KTH, School of Electrical Engineering and Computer Science (EECS), Computer Science, Computational Science and Technology (CST).ORCID iD: 0000-0003-4132-3175
KTH, School of Electrical Engineering and Computer Science (EECS), Intelligent systems, Decision and Control Systems (Automatic Control).ORCID iD: 0000-0002-9432-254x
KTH, School of Electrical Engineering and Computer Science (EECS), Intelligent systems, Decision and Control Systems (Automatic Control).ORCID iD: 0000-0002-4065-715x
(English)Manuscript (preprint) (Other academic)
Abstract [en]

Physics-Informed Neural Networks (PINNs) are a powerful deep learning method capable of providing solutions and parameter estimations of physical systems. Given the complexity of their neural network structure, the convergence speed is still limited compared to numerical methods, mainly when used in applications that model realistic systems. The network initialization follows a random distribution of the initial weights, as in the case of traditional neural networks, which could lead to severe model convergence bottlenecks. To overcome this problem, we follow current studies that deal with optimal initial weights in traditional neural networks. In this paper, we use a convex optimization model to improve the initialization of the weights in PINNs and accelerate convergence. We investigate two optimization models as a first training step, defined as pre-training, one involving only the boundaries and one including physics. The optimization is focused on the first layer of the neural network part of the PINN model, while the other weights are randomly initialized. We test the methods using a practical application of the heat diffusion equation to model the temperature distribution of power transformers. The PINN model with boundary pre-training is the fastest converging method at the current stage.

National Category
Artificial Intelligence
Identifiers
URN: urn:nbn:se:kth:diva-362872OAI: oai:DiVA.org:kth-362872DiVA, id: diva2:1955088
Funder
Vinnova, 2023-00241Vinnova, 2021-03748
Note

QC 20250430

Available from: 2025-04-28 Created: 2025-04-28 Last updated: 2025-05-02Bibliographically approved
In thesis
1. Scientific Machine Learning for Forward and Inverse Problems: Physics-Informed Neural Networks and Machine Learning Algorithms with Applications to Dynamical Systems
Open this publication in new window or tab >>Scientific Machine Learning for Forward and Inverse Problems: Physics-Informed Neural Networks and Machine Learning Algorithms with Applications to Dynamical Systems
2025 (English)Doctoral thesis, comprehensive summary (Other academic)
Abstract [en]

Scientific Machine Learning (SciML) is a promising field that combines data-driven models with physical laws and principles. A novel example is the application of Artificial Neural Networks (ANNs) to solve Ordinary Differential Equations (ODEs) and Partial Differential Equations (PDEs). One of the most recent approaches in this area is Physics-Informed Neural Networks (PINNs), which encode the governing physical equations directly into the neural network architecture. PINNs can solve both forward and inverse problems, learning the solution to differential equations and inferring unknown parameters or even functional forms. Therefore, they are particularly effective when partially known equations or incomplete models describe real-world systems. 

Differential equations enable a mathematical formulation for various fundamental physical laws. ODEs and PDEs are used to model the behavior of complex and dynamical systems in many fields of science. However, many real-world problems are either too complex to solve exactly or involve equations that are not fully known. In these cases, we rely on numerical methods to approximate solutions. While these methods can be very accurate, they often are computationally expensive, especially for large, nonlinear, or high-dimensional problems. Therefore, exploring alternative approaches like SciML to find more efficient and scalable solutions is fundamental.

This thesis presents a series of applications of SciML methods in identifying and solving real-world systems. First, we demonstrate using PINNs combined with symbolic regression to recover governing equations from sparse observational data, focusing on cellulose degradation within power transformers. PINNs are then applied to solve forward problems, specifically the 1D and 2D heat diffusion equations, which model thermal distribution in transformers. Moreover, we also develop an approach for optimal sensor placement using PINNs that improves data collection efficiency. A third case study examines how dimensionality reduction techniques, such as Principal Component Analysis (PCA), can be applied to explain and visualize high-dimensional data, where each observation comprises a large number of variables that describe physical systems. Using datasets on Cellulose Nanofibrils (CNFs) of various materials and concentrations, Machine Learning (ML) techniques are employed to characterize and interpret the system behavior. 

The second part of this thesis focuses on improving the scalability and robustness of PINNs. We propose a pretraining strategy that optimizes the initial weights, reducing stochasticity variability to address training instability and high computational costs in higher-dimensional problems arising from solving multi-dimensional or parametric PDEs. Moreover, we introduce an extension of PINNs, referred to as $PINN, which includes Bayesian probability within a domain decomposition framework. This formulation enhances performance, particularly in handling noisy data and multi-scale problems.

Abstract [sv]

Scientific Machine Learning (SciML) är ett lovande område som kombinerar datadrivna modeller med fysiska lagar och principer. Ett nytt exempel är tillämpningen av artificiella neurala nätverk (ANN) för att lösa ordinära differentialekvationer (ODE) och partiella differentialekvationer (PDE). Ett av de senaste tillvägagångssätten inom detta område är PINN (Physics-Informed Neural Networks), som kodar de styrande fysikaliska ekvationerna direkt i neuronnätets arkitektur. PINN kan lösa både framåtriktade och inversa problem, lära sig lösningen på differentialekvationer och härleda okända parametrar eller till och med funktions former. Därför är de särskilt effektiva när delvis kända ekvationer eller ofullständiga modeller beskriver verkliga system.

Differentialekvationer möjliggör en matematisk formulering av grundläggande fysiska lagar. ODE:er och PDE:er används för att modellera beteendet hos komplexa och dynamiska system inom många vetenskapliga områden. I verkligheten är många problem antingen för komplexa för att kunna lösas exakt eller innehåller ekvationer som inte är helt kända. I dessa fall förlitar vi oss på numeriska metoder för att approximera lösningar. Trots att dessa metoder kan vara mycket exakta är de ofta beräkningsmässigt dyra, särskilt för stora, olinjära eller mångdimensionella problem. Det är därför viktigt att utforska alternativa metoder som SciML för att hitta effektivare och mer skalbara lösningar.

I denna avhandling presenteras en serie tillämpningar av SciML-metoder för att identifiera och lösa verkliga system. Först demonstrerar vi hur PINN kombinerat med symbolisk regression kan användas för att återskapa styrande ekvationer från gles observationsdata, med fokus på cellulosanedbrytning i krafttransformatorer. PINNs används sedan för att lösa framåtriktade problem, särskilt 1D- och 2D-värmediffusionsekvationerna, som modellerar termisk distribution i transformatorer. Dessutom utvecklar vi ett tillvägagångssätt för optimal sensorplacering med hjälp av PINN som förbättrar datainsamlingseffektiviteten. I ett tredje användnings område undersöks hur tekniker för dimensionsreduktion, såsom Principal Component Analysis (PCA), kan tillämpas för att förklara och visualisera mångdimensionell data, där varje observation består av ett stort antal variabler som beskriver fysiska system. Med hjälp av dataset om cellulosa nanofibrillärer CNF) av olika material och koncentrationer används maskininlärningstekniker (ML) för att karakterisera och tolka systemets beteende.

Den andra delen av avhandlingen fokuserar på att förbättra skalbarheten och robustheten hos PINN. Vi föreslår en strategi för förträning som optimerar de initiala vikterna, vilket minskar stokasticitetsvariabiliteten för att hantera träningsinstabilitet och höga beräkningskostnader i problem med fler dimensioner som uppstår vid lösning av mångdimensionella eller parametriska PDE:er. Dessutom introducerar vi en tillägg för PINN, kallad $PINN, som inkluderar Bayesiansk sannolikhet inom ett ramverk för domänkomposition. Denna formulering förbättrar prestandan, särskilt vid hantering av brusiga data och flerskaliga problem.

Place, publisher, year, edition, pages
Stockholm: KTH Royal Institute of Technology, 2025. p. xii, 124
Series
TRITA-EECS-AVL ; 2025:44
Keywords
Scientific Machine Learning, Physics-Informed Neural Networks, System Identification, Inverse Problem, Forward Problem
National Category
Computer Sciences
Research subject
Computer Science
Identifiers
urn:nbn:se:kth:diva-363009 (URN)978-91-8106-263-2 (ISBN)
Public defence
2025-05-26, https://kth-se.zoom.us/j/66482272586, Kollegiesalen, Brinellvägen 6, Stockholm, 13:00 (English)
Opponent
Supervisors
Funder
Vinnova, 2021-03748
Note

QC 20250505

Available from: 2025-05-05 Created: 2025-05-02 Last updated: 2025-05-05Bibliographically approved

Open Access in DiVA

fulltext(630 kB)28 downloads
File information
File name FULLTEXT01.pdfFile size 630 kBChecksum SHA-512
fd827fcf69706f7dbd830dc5b47d4b3d341237eb86a5fd6ac490fa92fd50040a48c6b1b6b1e5be0464f7e94092288d75903f7834ba75ca949f8a52081772c25d
Type fulltextMimetype application/pdf

Search in DiVA

By author/editor
Li, SiruiBragone, FedericaBarreau, MatthieuMorozovska, Kateryna
By organisation
Decision and Control Systems (Automatic Control)Computational Science and Technology (CST)
Artificial Intelligence

Search outside of DiVA

GoogleGoogle Scholar
Total: 29 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

urn-nbn

Altmetric score

urn-nbn
Total: 285 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf