The conceptual design of liquid rocket propulsion systems is a fundamental but computationally intensive process, reliant on legacy physics-based software that creates a bottleneck to rapid design-space exploration. This thesis investigates the application of Artificial Intelligence (AI) to develop high-speed surrogate models as a modern alternative to these traditional tools. The primary objective is to create and validate neural network architectures to determine the feasibility of replacing traditional software with AI surrogates. The methodology involves generating a 14.5 million-point dataset from a legacy MATLAB program, which is then preprocessed using an Isolation Forest algorithm to remove statistical outliers. Three distinct deep learning architectures are systematically compared: a baseline Fully Connected Network (FCN), a deep Residual Network (ResNet) and a modern Transformer, whose hyperparameters are refined through an automated optimization campaign using the Optuna framework. The results prove that all surrogate models achieve a computational speed-up of over four orders of magnitude compared to the legacy code, and a detailed comparative analysis of predictive accuracy reveals that the ResNet is the superior architecture, consistently achieving the lowest error metrics across all output parameters. While the Transformer was less accurate, further analysis showed it had a unique robustness, naturally avoiding the unphysical outlier predictions generated by the other models. The study concludes that ResNet provides the best overall balance of predictive accuracy and computational efficiency, making it a viable replacement for the traditional simulation tool. These promising results validate the development of deep learning-based surrogates as an effective strategy, and this approach warrants extension in future studies to accelerate the engineering design and optimization process.
La progettazione concettuale dei sistemi a propulsione liquida è un processo fondamentale ma computazionalmente intensivo, basato su software tradizionali che limitano una rapida esplorazione dello spazio di progettazione: questa tesi analizza l'applicazione dell'Intelligenza Artificiale (IA) allo sviluppo di modelli surrogati ad alta velocità come loro alternativa. L'obiettivo è creare e validare architetture di reti neurali per determinare la fattibilità della sostituzione del software tradizionale con surrogati basati sull'IA. Il processo prevede la generazione, attraverso un codice MATLAB tradizionale, di un dataset di 14.5 milioni di elementi, che viene poi pre-processato con l'algoritmo Isolation Forest per rimuovere gli outlier statistici. Vengono confrontate parallelamente tre diverse architetture di deep learning: un semplice Fully Connected (FCN), una rete neurale residua (ResNet) e un Transformer, i cui iperparametri sono stati ottimizzati tramite un processo automatizzato con Optuna. I risultati dimostrano che tutti i modelli surrogati sono almeno quattro ordini di grandezza più veloci rispetto al codice tradizionale, e un'analisi comparativa della loro accuratezza predittiva prova che la ResNet è il surrogato migliore, con gli errori più bassi su tutti i parametri di output. Sebbene il Transformer sia risultato meno accurato, ha mostrato una robustezza unica, non fornendo predizioni anomale o fisicamente non plausibili, generate dagli altri modelli. Lo studio conclude che la ResNet offre il miglior bilanciamento complessivo tra accuratezza predittiva ed efficienza computazionale, rendendola un valido sostituto per lo strumento di simulazione tradizionale. Questi risultati convalidano l’efficacia della strategia intrapresa, e l’approccio merita quindi di essere approfondito in studi futuri, al fine di accelerare il processo di progettazione e ottimizzazione ingegneristica.
Development of high-fidelity surrogate models for the design of pressure-fed liquid rocket upper stages using neural networks
Separovic, Tomislav Marko;Conti, Matteo
2024/2025
Abstract
The conceptual design of liquid rocket propulsion systems is a fundamental but computationally intensive process, reliant on legacy physics-based software that creates a bottleneck to rapid design-space exploration. This thesis investigates the application of Artificial Intelligence (AI) to develop high-speed surrogate models as a modern alternative to these traditional tools. The primary objective is to create and validate neural network architectures to determine the feasibility of replacing traditional software with AI surrogates. The methodology involves generating a 14.5 million-point dataset from a legacy MATLAB program, which is then preprocessed using an Isolation Forest algorithm to remove statistical outliers. Three distinct deep learning architectures are systematically compared: a baseline Fully Connected Network (FCN), a deep Residual Network (ResNet) and a modern Transformer, whose hyperparameters are refined through an automated optimization campaign using the Optuna framework. The results prove that all surrogate models achieve a computational speed-up of over four orders of magnitude compared to the legacy code, and a detailed comparative analysis of predictive accuracy reveals that the ResNet is the superior architecture, consistently achieving the lowest error metrics across all output parameters. While the Transformer was less accurate, further analysis showed it had a unique robustness, naturally avoiding the unphysical outlier predictions generated by the other models. The study concludes that ResNet provides the best overall balance of predictive accuracy and computational efficiency, making it a viable replacement for the traditional simulation tool. These promising results validate the development of deep learning-based surrogates as an effective strategy, and this approach warrants extension in future studies to accelerate the engineering design and optimization process.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_11_Conti_Separovic_Thesis.pdf
non accessibile
Descrizione: Thesis Document
Dimensione
40.8 MB
Formato
Adobe PDF
|
40.8 MB | Adobe PDF | Visualizza/Apri |
|
2025_11_Conti_Separovic_Executive_Summary.pdf
non accessibile
Descrizione: Executive Summary Document
Dimensione
1.49 MB
Formato
Adobe PDF
|
1.49 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/10589/246732