This thesis explores a novel data-driven approach for simulating partial differential equations (PDEs), which are crucial for understanding complex phenomena in physics, biology, and engineering. Traditional numerical methods often struggle with the PDEs complexity, and machine learning techniques open new avenues for both discovery and prediction of dynamical systems based on empirical data. Given the challenges of obtaining dense measurements across physical domains, this work introduces a new deep learning-based strategy, named {\em parametric SINDy-SHRED}, which exploits sparse sensor measurements to model and predict the dynamic of high-dimensional state variables. The method combines Sparse Identification of Non-linear Dynamics (SINDy) with a SHallow REcurrent Decoder (SHRED) network. Specifically, SINDy extracts a compact and interpretable dynamical system from the temporal history of sparse sensor data, which are encoded by SHRED through a recurrent neural network. The system identified by SINDy includes parametric dependencies, enhancing the understanding of the system's underlying mechanisms. Moreover, SHRED projects the identified latent state at the reduced level onto the original high-dimensional state space through a shallow decoder. Parametric SINDy-SHRED enables {\em (i)} to predict spatio-temporal evolutions of high-dimensional dynamical systems starting from limited sensor data and {\em (ii)} to identify a sparse and compact dynamical system at the latent level, thus enhancing robustness and reliability. The effectiveness of the proposed methodology is validated through numerical results from various test cases, demonstrating its potential to enhance scientific modelling and predictive analytics. This approach not only enable accurate forecasts but also deepens the understanding of complex dynamical systems, bridging possible gaps between theoretical models and practical applications.
Questa tesi esplora un nuovo approccio basato sui dati per l'analisi delle equazioni differenziali parziali (EDP), cruciali per comprendere fenomeni complessi in fisica, biologia e ingegneria. I metodi analitici tradizionali spesso faticano a gestire la complessità delle EDP, e la disponibilità del machine learning apre nuove vie sia per la scoperta sia per la previsione della dinamica dei sistemi basata su dati empirici. Data la difficoltà di ottenere misurazioni dense attraverso domini fisici, questo lavoro introduce una strategia ibrida di rete neurale, parametric SINDy-SHRED, che utilizza efficacemente dati sparsi da sensori limitati. Il metodo combina l'Identificazione Sparsa di Dinamiche Non-lineari (SINDy) con una rete di Decoder Ricorrente Superficiale (SHRED). SINDy estrae dinamiche chiare e interpretabili da dati sparsi, migliorando la comprensione dei meccanismi sottostanti del sistema includendo dipendenze parametriche. Contemporaneamente, SHRED utilizza un'unità ricorrente per la filtrazione del rumore e un decoder per proiettare input ridotti dimensionalmente di nuovo a ricostruzioni spaziali complete. Questa integrazione, e l'uso di parametri di controllo, permette al modello non solo di prevedere l'evoluzione spazio-temporale completa di un sistema con dati limitati ma anche di identificare le equazioni governative, migliorando così la robustezza e l'affidabilità del metodo. L'efficacia della metodologia proposta è validata attraverso risultati numerici da vari scenari di test, dimostrando il suo potenziale per migliorare la modellazione scientifica e l'analitica predittiva. Questo approccio facilita previsioni accurate e approfondisce la comprensione dei sistemi dinamici complessi, colmando il divario tra modelli teorici e applicazioni pratiche.
Parametric SINDy-SHRED: a data-driven deep learning-based approach to predict and discover parametrical dynamical systems
COLETTI, GREGORIO
2024/2025
Abstract
This thesis explores a novel data-driven approach for simulating partial differential equations (PDEs), which are crucial for understanding complex phenomena in physics, biology, and engineering. Traditional numerical methods often struggle with the PDEs complexity, and machine learning techniques open new avenues for both discovery and prediction of dynamical systems based on empirical data. Given the challenges of obtaining dense measurements across physical domains, this work introduces a new deep learning-based strategy, named {\em parametric SINDy-SHRED}, which exploits sparse sensor measurements to model and predict the dynamic of high-dimensional state variables. The method combines Sparse Identification of Non-linear Dynamics (SINDy) with a SHallow REcurrent Decoder (SHRED) network. Specifically, SINDy extracts a compact and interpretable dynamical system from the temporal history of sparse sensor data, which are encoded by SHRED through a recurrent neural network. The system identified by SINDy includes parametric dependencies, enhancing the understanding of the system's underlying mechanisms. Moreover, SHRED projects the identified latent state at the reduced level onto the original high-dimensional state space through a shallow decoder. Parametric SINDy-SHRED enables {\em (i)} to predict spatio-temporal evolutions of high-dimensional dynamical systems starting from limited sensor data and {\em (ii)} to identify a sparse and compact dynamical system at the latent level, thus enhancing robustness and reliability. The effectiveness of the proposed methodology is validated through numerical results from various test cases, demonstrating its potential to enhance scientific modelling and predictive analytics. This approach not only enable accurate forecasts but also deepens the understanding of complex dynamical systems, bridging possible gaps between theoretical models and practical applications.File | Dimensione | Formato | |
---|---|---|---|
TESI_parametricSindyShred_GregorioColetti.pdf
accessibile in internet per tutti a partire dal 13/03/2026
Descrizione: Tesi di ricerca in formato articolo
Dimensione
4.74 MB
Formato
Adobe PDF
|
4.74 MB | Adobe PDF | Visualizza/Apri |
Executive_Summary_Gregorio_Coletti.pdf
accessibile in internet per tutti a partire dal 13/03/2026
Descrizione: Executive Summary
Dimensione
963.61 kB
Formato
Adobe PDF
|
963.61 kB | 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/10589/235270