This work explores the sequential-in-time framework for solving time-dependent Partial Differential Equations (PDEs) using nonlinear parametrizations such as neural networks. In contrast to conventional methods that represent the entire spatio-temporal solution within a single model, the sequential-in-time approach approximates the solution through a nonlinear parametrization at each time instant, dynamically updating the parameters over time. Focusing on linear combination of parameter-dependent basis functions, we introduce two novel shallow neural network architectures tailored to this framework: the Shaped Radial Basis Function Network (SRBFN) and the Piecewise Polynomial-Like Network (PwPLN). The SRBFN extends traditional Radial Basis Function Networks by incorporating a shape parameter in each basis function, enhancing its flexibility in function approximation, while the PwPLN modifies piecewise polynomial functions, making the knot positions trainable parameters, enabling them to adapt dynamically. Additionally, we integrate an active learning approach within the SRBFN, guiding adaptive sampling in regions of interest based on the basis functions and their derivatives. We evaluate the proposed architectures on two electrophysiology test cases—the Simplified Nagumo and Mitchell-Schaeffer equations—demonstrating the effectiveness of these networks in capturing complex, time-dependent dynamics.
Questo lavoro esplora il framework sequential-in-time per la risoluzione di Equazioni Differenziali alle Derivate Parziali (EDP) dipendenti dal tempo utilizzando parametrizzazioni non lineari come le reti neurali. A differenza dei metodi convenzionali che rappresentano l'intera soluzione spazio-temporale all'interno di un singolo modello, l'approccio sequential-in-time approssima la soluzione attraverso una parametrizzazione non lineare in ogni istante di tempo, aggiornando dinamicamente i parametri nel tempo. Concentrandosi sulla combinazione lineare di funzioni base dipendenti dai parametri, introduciamo due nuove architetture di reti neurali superficiali per questo framework: la Shaped Radial Basis Function Network (SRBFN) e la Piecewise Polynomial-Like Network (PwPLN). La SRBFN estende Radial Basis Function Networks tradizionali incorporando un parametro di forma in ogni funzione base, migliorando la flessibilità nell’approssimare funzioni, mentre la PwPLN modifica le funzioni polinomiali a tratti, rendendo le posizioni dei nodi parametri addestrabili, permettendo loro di adattarsi dinamicamente. Inoltre, integriamo l'active learning sulla SRBFN, guidando il campionamento nelle regioni di interesse basato sulle funzioni base e sulle loro derivate. Valutiamo le architetture proposte su due casi di test in elettrofisiologia—le equazioni semplificate di Nagumo e di Mitchell-Schaeffer—dimostrando l'efficacia di queste reti nel catturare dinamiche complesse e dipendenti dal tempo.
Sequential in time learning of nonlinear parametrizations for PDE solutions in electrophysiology
Morales Lopez, Oswaldo Jesus
2023/2024
Abstract
This work explores the sequential-in-time framework for solving time-dependent Partial Differential Equations (PDEs) using nonlinear parametrizations such as neural networks. In contrast to conventional methods that represent the entire spatio-temporal solution within a single model, the sequential-in-time approach approximates the solution through a nonlinear parametrization at each time instant, dynamically updating the parameters over time. Focusing on linear combination of parameter-dependent basis functions, we introduce two novel shallow neural network architectures tailored to this framework: the Shaped Radial Basis Function Network (SRBFN) and the Piecewise Polynomial-Like Network (PwPLN). The SRBFN extends traditional Radial Basis Function Networks by incorporating a shape parameter in each basis function, enhancing its flexibility in function approximation, while the PwPLN modifies piecewise polynomial functions, making the knot positions trainable parameters, enabling them to adapt dynamically. Additionally, we integrate an active learning approach within the SRBFN, guiding adaptive sampling in regions of interest based on the basis functions and their derivatives. We evaluate the proposed architectures on two electrophysiology test cases—the Simplified Nagumo and Mitchell-Schaeffer equations—demonstrating the effectiveness of these networks in capturing complex, time-dependent dynamics.File | Dimensione | Formato | |
---|---|---|---|
2024_12_Morales_Executive Summary.pdf
non accessibile
Descrizione: Executive Summary
Dimensione
915.51 kB
Formato
Adobe PDF
|
915.51 kB | Adobe PDF | Visualizza/Apri |
2024_12_Morales.pdf
non accessibile
Descrizione: Thesis
Dimensione
2.43 MB
Formato
Adobe PDF
|
2.43 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/231144