The optimization of chemical process flowsheets based on rigorous first-principles simulations is often computationally demanding due to strong nonlinearities, recycle loops, and repeated convergence procedures, which can limit the practical applicability of simulation-based optimization in industrial contexts. This thesis develops and validates a fully automated and customizable surrogate-based hybrid modeling framework for optimization within the Aspen HYSYS environment. The proposed methodology generates structured datasets from rigorous simulations through space-filling Design of Experiments, constructs explicit polynomial surrogate models using automated regression procedures, and integrates them into the simulator via the User Model interface. The framework is demonstrated on a methanol synthesis process considering multiple surrogate-based hybrid configurations and optimization scenarios involving both operating and design variables. Several optimization algorithms are systematically evaluated to assess the impact of surrogate integration on convergence behavior, computational effort, constraint satisfaction, and economic performance. Surrogate-based hybrid formulations improve numerical robustness and reduce computational time when recycle-intensive sections are replaced, with CPU time reductions exceeding one order of magnitude in representative scenarios. Rigorous post-optimization validation shows that predictive accuracy is preserved within the training domain, with average validation errors of approximately 3–4% for methanol production and 2–3% for operating cost. Hybrid configurations based on two independent surrogate models for synthesis and separation provide the most favorable balance between predictive accuracy and numerical stability, with average validation errors on methanol production in the range of 1–2% and typically around 2–3% on operating expenditures (OpEx), together with consistent post-validation constraint satisfaction for the most robust optimization algorithms considered. The proposed framework provides a transparent and reproducible alternative to modeling tools integrated in commercial process simulators, offering greater freedom, flexibility, and control in the development of hybrid models and simulation-based optimization of complex chemical processes in industrial steady-state environments.
L’ottimizzazione di schemi di processo basata su simulazioni rigorose di tipo first-principles è spesso computazionalmente onerosa a causa di forti non linearità, ricicli e procedure iterative di convergenza ripetute, che possono limitarne l’applicabilità pratica in contesti industriali. Questa tesi sviluppa e valida un framework completamente automatizzato e personalizzabile di modellazione ibrida basata su modelli surrogati per l’ottimizzazione all’interno dell’ambiente di simulazione Aspen HYSYS. La metodologia proposta genera dataset strutturati a partire da simulazioni rigorose mediante Design of Experiments space-filling, costruisce modelli surrogati polinomiali espliciti tramite procedure di regressione automatizzate e li integra nel simulatore attraverso l’interfaccia User Model. Il framework è dimostrato su un processo di sintesi del metanolo considerando diverse configurazioni ibride basate su modelli surrogati e scenari di ottimizzazione che includono variabili operative e di progetto. Diversi algoritmi di ottimizzazione sono valutati sistematicamente per analizzare l’effetto dell’integrazione dei modelli surrogati sulla convergenza, sul costo computazionale, sulla soddisfazione dei vincoli e sulle prestazioni economiche. Le configurazioni ibride basate su modelli surrogati migliorano la robustezza numerica e riducono il tempo di calcolo quando vengono sostituite le sezioni del processo caratterizzate da strutture di riciclo complesse, con riduzioni del tempo di esecuzione superiori a un ordine di grandezza nei casi analizzati. La validazione rigorosa post-ottimizzazione conferma che l’accuratezza predittiva è mantenuta all’interno del dominio di addestramento, con errori medi di circa 3–4% sulla produzione di metanolo e 2–3% sui costi operativi (OpEx). Le configurazioni ibride basate su due modelli surrogati indipendenti per sintesi e separazione offrono il miglior compromesso tra accuratezza e stabilità numerica, con errori medi di validazione dell’ordine dell’1–2% sulla produzione di metanolo e mediamente intorno al 2–3% sull’OpEx, oltre a una soddisfazione dei vincoli confermata dopo la validazione rigorosa per gli algoritmi più robusti considerati. Il framework proposto offre un’alternativa trasparente e riproducibile agli strumenti di modellazione integrati nei simulatori di processo commerciali, garantendo maggiore libertà, flessibilità e controllo nella costruzione di modelli ibridi e nelle attività di ottimizzazione basata su simulazione di processi chimici complessi in ambienti industriali a regime stazionario.
Automated surrogate-based hybrid modeling for simulation-based optimization in Aspen HYSYS
Grimoldi, Marco
2024/2025
Abstract
The optimization of chemical process flowsheets based on rigorous first-principles simulations is often computationally demanding due to strong nonlinearities, recycle loops, and repeated convergence procedures, which can limit the practical applicability of simulation-based optimization in industrial contexts. This thesis develops and validates a fully automated and customizable surrogate-based hybrid modeling framework for optimization within the Aspen HYSYS environment. The proposed methodology generates structured datasets from rigorous simulations through space-filling Design of Experiments, constructs explicit polynomial surrogate models using automated regression procedures, and integrates them into the simulator via the User Model interface. The framework is demonstrated on a methanol synthesis process considering multiple surrogate-based hybrid configurations and optimization scenarios involving both operating and design variables. Several optimization algorithms are systematically evaluated to assess the impact of surrogate integration on convergence behavior, computational effort, constraint satisfaction, and economic performance. Surrogate-based hybrid formulations improve numerical robustness and reduce computational time when recycle-intensive sections are replaced, with CPU time reductions exceeding one order of magnitude in representative scenarios. Rigorous post-optimization validation shows that predictive accuracy is preserved within the training domain, with average validation errors of approximately 3–4% for methanol production and 2–3% for operating cost. Hybrid configurations based on two independent surrogate models for synthesis and separation provide the most favorable balance between predictive accuracy and numerical stability, with average validation errors on methanol production in the range of 1–2% and typically around 2–3% on operating expenditures (OpEx), together with consistent post-validation constraint satisfaction for the most robust optimization algorithms considered. The proposed framework provides a transparent and reproducible alternative to modeling tools integrated in commercial process simulators, offering greater freedom, flexibility, and control in the development of hybrid models and simulation-based optimization of complex chemical processes in industrial steady-state environments.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_3_Grimoldi_Thesis.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Tesi
Dimensione
1.19 MB
Formato
Adobe PDF
|
1.19 MB | Adobe PDF | Visualizza/Apri |
|
2026_3_Grimoldi_Executive_Summary.pdf
accessibile in internet solo dagli utenti autorizzati
Descrizione: Executive Summary
Dimensione
574.08 kB
Formato
Adobe PDF
|
574.08 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/10589/253143