This PhD thesis investigates how Machine Learning (ML) can support design for environmental sustainability by integrating data-driven models with Life Cycle Assessment (LCA). While LCA is a consolidated methodology for quantifying environmental impacts across a product’s life cycle, from raw material extraction to end-of-life management, its traditional application is often time-consuming and poorly suited to the conceptual design phase, where information is limited but decisions are highly influential. The work addresses this gap by developing ML-based surrogate models capable of rapidly estimating environmental and process-related parameters, thereby enabling designers to consider sustainability alongside performance, cost, and reliability from the earliest stages of design. The research adopts and adapts the CRISP-DM (Cross-Industry Standard Process for Data Mining) methodology as a general framework to structure the entire workflow: Business Understanding, Data Understanding, Data Preparation, Modeling, Evaluation, and Deployment. Within this framework, specific attention is devoted to defining acceptable model accuracy, selecting appropriate error metrics and addressing uncertainty in real industrial data. Data-related activities, such as system boundary definition, configuration parameter extraction from CAD/PDM systems, outlier detection via interquartile range analysis, data augmentation through geometric scaling and scenario definition, and feature selection, are treated as central enablers of model robustness and interpretability. The methodology is validated through three main themes, each articulated into one or more case studies. The first concerns the design of electric vehicles at the assembly level, with cradle-to-grave system boundaries. Here, ML models are trained on LCA-derived datasets and used to predict specific environmental impact, supporting sustainability-oriented design choices at the system scale. The second theme focuses on manufactured components (cradle-to-gate), including a steam turbine blade carrier and a mechanical shaft. Analytical models and multiple ML algorithms (e.g., linear regression, decision trees, random forests, gradient boosted trees, and neural networks) are compared for predicting process parameters and environmental indicators, highlighting the role of data augmentation and feature selection in improving accuracy and sensitivity to design variations. The third theme addresses the end-of-life stage and the disassembly of rusted mechanical joints. Experimental campaigns support the development of ML models for predicting disassembly time and, in a further extension, a Convolutional Neural Network (CNN) for classifying rust degree from images, demonstrating the applicability of the method to both numerical regression and image-based classification. Across all applications, model interpretability is emphasized through feature importance analysis, enabling designers to understand which geometric and process parameters most significantly influence environmental impacts and operational performance. The resulting models are conceived for integration into existing IT infrastructures (e.g., CAD-based tools, web services), making them accessible to design teams without requiring advanced expertise in LCA or AI. Overall, the thesis shows that combining ML with LCA within a CRISP-DM-based framework can transform environmental assessment from a late verification step into a proactive decision-support tool. The proposed approach supports faster, more informed, and more sustainable design decisions, and it is scalable to different objectives, including cost-oriented design. This work therefore contributes a general, flexible, and interpretable methodology for embedding artificial intelligence into industrial design processes to foster sustainability and circularity.
Questa tesi di dottorato indaga come il Machine Learning (ML) possa supportare la progettazione per la sostenibilità ambientale integrando modelli basati sui dati con la Life Cycle Assessment (LCA). Sebbene la LCA sia una metodologia consolidata per quantificare gli impatti ambientali lungo l’intero ciclo di vita di un prodotto, dall’estrazione delle materie prime alla gestione del fine vita, la sua applicazione tradizionale risulta spesso dispendiosa in termini di tempo e poco adatta alla fase di progettazione concettuale, in cui le informazioni disponibili sono limitate ma le decisioni hanno un’influenza elevata. Il lavoro affronta questo divario sviluppando modelli surrogati basati su ML, in grado di stimare rapidamente parametri ambientali e di processo, consentendo così ai progettisti di considerare la sostenibilità insieme a prestazioni, costi e affidabilità fin dalle primissime fasi della progettazione. La ricerca adotta e adatta la metodologia CRISP-DM (Cross-Industry Standard Process for Data Mining) come quadro di riferimento generale per strutturare l’intero flusso di lavoro: Business Understanding, Data Understanding, Data Preparation, Modeling, Evaluation e Deployment. All’interno di questo framework, viene dedicata particolare attenzione alla definizione di livelli accettabili di accuratezza dei modelli, alla selezione di metriche di errore appropriate e alla gestione dell’incertezza nei dati industriali reali. Le attività legate ai dati, quali la definizione dei confini di sistema, l’estrazione dei parametri di configurazione dai sistemi CAD/PDM, l’individuazione degli outlier tramite analisi dell’intervallo interquartile, l’aumento dei dati mediante scalatura geometrica e definizione di scenari, nonché la selezione delle caratteristiche (feature selection), sono trattate come elementi centrali per garantire la robustezza e l’interpretabilità dei modelli. La metodologia è validata attraverso tre temi principali, ciascuno articolato in uno o più casi studio. Il primo riguarda la progettazione di veicoli elettrici a livello di assemblaggio, con confini di sistema cradle-to-grave. In questo contesto, i modelli di ML sono addestrati su dataset derivati da LCA e utilizzati per prevedere specifici impatti ambientali, supportando scelte progettuali orientate alla sostenibilità su scala di sistema. Il secondo tema si concentra su componenti manifatturieri (cradle-to-gate), tra cui un supporto per pala di turbina a vapore e un albero meccanico. Modelli analitici e diversi algoritmi di ML (ad esempio regressione lineare, alberi decisionali, random forest, alberi gradient boosted e reti neurali) sono confrontati per la previsione di parametri di processo e indicatori ambientali, evidenziando il ruolo dell’aumento dei dati e della selezione delle caratteristiche nel migliorare l’accuratezza e la sensibilità alle variazioni progettuali. Il terzo tema affronta la fase di fine vita e lo smontaggio di giunzioni meccaniche arrugginite. Campagne sperimentali supportano lo sviluppo di modelli di ML per la previsione del tempo di smontaggio e, in un’ulteriore estensione, di una Convolutional Neural Network (CNN) per la classificazione del grado di ruggine a partire da immagini, dimostrando l’applicabilità del metodo sia a problemi di regressione numerica sia di classificazione basata su immagini. In tutte le applicazioni, l’interpretabilità dei modelli è enfatizzata attraverso l’analisi dell’importanza delle caratteristiche, consentendo ai progettisti di comprendere quali parametri geometrici e di processo influenzino maggiormente gli impatti ambientali e le prestazioni operative. I modelli risultanti sono concepiti per l’integrazione nelle infrastrutture IT esistenti (ad esempio strumenti basati su CAD o servizi web), rendendoli accessibili ai team di progettazione senza richiedere competenze avanzate in LCA o intelligenza artificiale. Nel complesso, la tesi dimostra che la combinazione di ML e LCA all’interno di un framework basato su CRISP-DM può trasformare la valutazione ambientale da una fase di verifica tardiva a uno strumento proattivo di supporto alle decisioni. L’approccio proposto consente decisioni progettuali più rapide, informate e sostenibili ed è scalabile a diversi obiettivi, inclusa la progettazione orientata ai costi. Questo lavoro contribuisce pertanto con una metodologia generale, flessibile e interpretabile per integrare l’intelligenza artificiale nei processi di progettazione industriale, favorendo sostenibilità e circolarità.
Design for Environmental Sustainability through Machine Learning Based Method / Manuguerra, Luca. - (2026 Mar 31).
Design for Environmental Sustainability through Machine Learning Based Method
MANUGUERRA, LUCA
2026-03-31
Abstract
This PhD thesis investigates how Machine Learning (ML) can support design for environmental sustainability by integrating data-driven models with Life Cycle Assessment (LCA). While LCA is a consolidated methodology for quantifying environmental impacts across a product’s life cycle, from raw material extraction to end-of-life management, its traditional application is often time-consuming and poorly suited to the conceptual design phase, where information is limited but decisions are highly influential. The work addresses this gap by developing ML-based surrogate models capable of rapidly estimating environmental and process-related parameters, thereby enabling designers to consider sustainability alongside performance, cost, and reliability from the earliest stages of design. The research adopts and adapts the CRISP-DM (Cross-Industry Standard Process for Data Mining) methodology as a general framework to structure the entire workflow: Business Understanding, Data Understanding, Data Preparation, Modeling, Evaluation, and Deployment. Within this framework, specific attention is devoted to defining acceptable model accuracy, selecting appropriate error metrics and addressing uncertainty in real industrial data. Data-related activities, such as system boundary definition, configuration parameter extraction from CAD/PDM systems, outlier detection via interquartile range analysis, data augmentation through geometric scaling and scenario definition, and feature selection, are treated as central enablers of model robustness and interpretability. The methodology is validated through three main themes, each articulated into one or more case studies. The first concerns the design of electric vehicles at the assembly level, with cradle-to-grave system boundaries. Here, ML models are trained on LCA-derived datasets and used to predict specific environmental impact, supporting sustainability-oriented design choices at the system scale. The second theme focuses on manufactured components (cradle-to-gate), including a steam turbine blade carrier and a mechanical shaft. Analytical models and multiple ML algorithms (e.g., linear regression, decision trees, random forests, gradient boosted trees, and neural networks) are compared for predicting process parameters and environmental indicators, highlighting the role of data augmentation and feature selection in improving accuracy and sensitivity to design variations. The third theme addresses the end-of-life stage and the disassembly of rusted mechanical joints. Experimental campaigns support the development of ML models for predicting disassembly time and, in a further extension, a Convolutional Neural Network (CNN) for classifying rust degree from images, demonstrating the applicability of the method to both numerical regression and image-based classification. Across all applications, model interpretability is emphasized through feature importance analysis, enabling designers to understand which geometric and process parameters most significantly influence environmental impacts and operational performance. The resulting models are conceived for integration into existing IT infrastructures (e.g., CAD-based tools, web services), making them accessible to design teams without requiring advanced expertise in LCA or AI. Overall, the thesis shows that combining ML with LCA within a CRISP-DM-based framework can transform environmental assessment from a late verification step into a proactive decision-support tool. The proposed approach supports faster, more informed, and more sustainable design decisions, and it is scalable to different objectives, including cost-oriented design. This work therefore contributes a general, flexible, and interpretable methodology for embedding artificial intelligence into industrial design processes to foster sustainability and circularity.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


