In response to the growing need for sustainable Artificial Intelligence (AI), further accentuated by rising energy costs in Europe, this thesis investigates how three design levers, namely encoder architecture, input resolution, and structured pruning, jointly shape the trade-off between segmentation accuracy and energy efficiency in deep learning for Additive Manufacturing (AM). By systematically combining these factors across two controlled experiments and analyzing their behavior under realistic operating conditions that emulate typical constraints in domains such as defence, space, and emergency response, the study develops a framework to support energy-aware model selection and deployment in Laser Powder Bed Fusion (L-PBF). Motivated by the limits of accuracy-centric reporting, actual power usage is measured during training and inference with CodeCarbon under a fixed, replicable protocol, operationalizing Green AI in an industrial setting and demonstrating how concrete design choices can make AM segmentation genuinely greener. Empirical findings reveal that higher resolutions and heavier encoders deliver diminishing accuracy returns at sharply increasing energy costs, while lightweight architectures such as ResNet-18 achieve comparable performance at a fraction of the consumption. Building on this insight, structured pruning further enhances efficiency, preserving accuracy while reducing inference energy and latency by up to 70-75%. By bridging methodological rigor with industrial relevance, this work offers one of the first empirically grounded demonstrations of energy-aware deep learning in AM, establishes replicable reporting standards for future research, and proposes scenario-based guidelines for sustainable AI deployment. Beyond the specific domain of AM, the framework underscores that efficiency is not an intrinsic property of a model but an emergent outcome of its deployment context, representing a step toward AI systems that are not only intelligent, but also transparent and energy-conscious.
In risposta alla crescente attenzione verso un’Intelligenza Artificiale (IA) sostenibile, ulteriormente accentuata dall’aumento dei costi energetici in Europa, questa tesi indaga come tre leve progettuali, ovvero l’architettura dell’encoder, la risoluzione dell’input e il pruning strutturato, influenzino congiuntamente il compromesso tra accuratezza di segmentazione ed efficienza energetica nel deep learning applicato all’Additive Manufacturing (AM). Combinando sistematicamente questi fattori in due esperimenti controllati e analizzandone il comportamento in condizioni operative realistiche, che emulano i vincoli tipici di domini quali difesa, spazio e contesti di emergenza, questo studio sviluppa un framework a supporto della selezione e dell’utilizzo di modelli energeticamente efficienti nel processo di Laser Powder Bed Fusion (L-PBF). Spinta dai limiti di una valutazione puramente centrata sull’accuratezza, la ricerca misura il consumo energetico reale durante il training e l’inferenza mediante CodeCarbon, utilizzando un protocollo fisso e replicabile. In tal modo, la tesi traduce i principi del Green AI in un contesto industriale, dimostrando come scelte progettuali mirate possano rendere la segmentazione per AM concretamente più green. I risultati empirici mostrano che risoluzioni più elevate ed encoder più complessi forniscono incrementi marginali di accuratezza a fronte di costi energetici nettamente superiori, mentre architetture leggere come ResNet-18 raggiungono prestazioni comparabili con una frazione dei consumi. Sulla base di questa evidenza, il pruning strutturato migliora ulteriormente l’efficienza, preservando l’accuratezza e riducendo il consumo energetico e la latenza in inferenza fino al 70–75%. Unendo rigore metodologico e rilevanza industriale, questo lavoro fornisce una delle prime dimostrazioni empiriche di deep learning energeticamente consapevole in AM, definisce metriche replicabili per la ricerca futura e propone linee guida basate su scenari per un’implementazione sostenibile dell’IA. Il framework evidenzia, anche al di fuori del dominio specifico dell’AM, come l’efficienza non sia una proprietà intrinseca del modello, bensì un risultato emergente dal contesto di utilizzo, rappresentando un passo verso sistemi di IA non solo intelligenti, ma anche trasparenti ed energeticamente consapevoli.
Green AI in additive manufacturing: energy-aware deep learning for quality assessment of complex shapes
Bruni, Alessia;DE CATALDO, PAOLO
2024/2025
Abstract
In response to the growing need for sustainable Artificial Intelligence (AI), further accentuated by rising energy costs in Europe, this thesis investigates how three design levers, namely encoder architecture, input resolution, and structured pruning, jointly shape the trade-off between segmentation accuracy and energy efficiency in deep learning for Additive Manufacturing (AM). By systematically combining these factors across two controlled experiments and analyzing their behavior under realistic operating conditions that emulate typical constraints in domains such as defence, space, and emergency response, the study develops a framework to support energy-aware model selection and deployment in Laser Powder Bed Fusion (L-PBF). Motivated by the limits of accuracy-centric reporting, actual power usage is measured during training and inference with CodeCarbon under a fixed, replicable protocol, operationalizing Green AI in an industrial setting and demonstrating how concrete design choices can make AM segmentation genuinely greener. Empirical findings reveal that higher resolutions and heavier encoders deliver diminishing accuracy returns at sharply increasing energy costs, while lightweight architectures such as ResNet-18 achieve comparable performance at a fraction of the consumption. Building on this insight, structured pruning further enhances efficiency, preserving accuracy while reducing inference energy and latency by up to 70-75%. By bridging methodological rigor with industrial relevance, this work offers one of the first empirically grounded demonstrations of energy-aware deep learning in AM, establishes replicable reporting standards for future research, and proposes scenario-based guidelines for sustainable AI deployment. Beyond the specific domain of AM, the framework underscores that efficiency is not an intrinsic property of a model but an emergent outcome of its deployment context, representing a step toward AI systems that are not only intelligent, but also transparent and energy-conscious.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Bruni_De Cataldo_Executive Summary.pdf
accessibile in internet per tutti
Descrizione: Testo executive summary
Dimensione
663.7 kB
Formato
Adobe PDF
|
663.7 kB | Adobe PDF | Visualizza/Apri |
|
2025_12_Bruni_De Cataldo_Tesi.pdf
accessibile in internet per tutti
Descrizione: Testo tesi
Dimensione
4.01 MB
Formato
Adobe PDF
|
4.01 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/10589/246570