In the design of a public work, cost estimation is the phase on which the economic credibility of the project depends, since the bill of quantities and cost estimate (computo metrico estimativo) determines the base amount put out to tender. The Italian Public Procurement Code (Legislative Decree 36/2023) provides for the use of digital information management methods and tools for construction, as well as for the use of regional price lists to determine the amounts of the works items. Currently, widely used estimating tools do not allow the objects of the information model to be linked directly to price list items; the association remains largely manual, because it requires two semantic structures designed for different purposes to be made to communicate. The thesis aims to define and validate a semi-automatic workflow, based on a Large Language Model (LLM), that assists the quantity surveyor in this association, and to evaluate the results against a set of metrics. The workflow is structured in three phases. In the first, the objects and their properties are extracted from the IFC file; in the second, the price list, organised in a database, is queried to identify candidate items for each object, and the language model ranks them according to the extracted properties; in the third, after the user has confirmed the associations, quantities and costs are calculated deterministically. Artificial intelligence (AI) is entrusted neither with the calculation of quantities nor with the writing of codes and prices, which remain bound to the information model and to the price list. The validation was carried out on the structure of a four-level building, exported in three IFC configurations that differ in the information content of the objects. The estimates produced by the workflow were compared with a manual reference estimate prepared on the Emilia-Romagna 2025 Price List, through four families of indicators relating to the correctness of the associations between objects and items, the deviation in quantities, the deviation in cost, and efficiency in terms of processing time and cost. Four iterations were performed for each configuration; these correspond to successive versions of the workflow and not to independent repetitions of the same test. In the final iteration, the deviation in total cost from the reference is below 1% in all three configurations, with preparation times ranging from 76 to 134 minutes, compared with the approximately six hours required for the manual estimate. Moreover, the aggregate figure should be read together with the deviations on individual items, which are larger and partly offset each other in the total. The workflow does not deliver a definitive estimate but a starting point for the professional, who is relieved of repetitive steps and of the preliminary check normally carried out by the operator. The experimentation has some limitations: the first is that the case study is a small building covering the structural discipline only; the second is that the workflow depends on a subscription that enables its use. A possible development is the use of models run on local infrastructure, which would guarantee the confidentiality of project data and the absence of subscription costs, albeit with a possible reduction in the accuracy of the results.
Nella progettazione di un'opera pubblica la stima dei costi è la fase da cui dipende la credibilità economica del progetto, poiché il computo metrico estimativo determina l'importo posto a base di gara. Il Codice dei contratti pubblici (D.Lgs. 36/2023) prevede l'uso di metodi e strumenti di gestione informativa digitale delle costruzioni e il ricorso ai prezzari regionali per la determinazione degli importi delle lavorazioni. Oggigiorno, gli strumenti di computo diffusi non consentono di collegare in modo diretto gli oggetti del modello informativo alle voci del prezzario, l'associazione resta in larga parte manuale, perché richiede di far dialogare due strutture semantiche pensate per scopi diversi. La tesi si propone di definire e validare un flusso di lavoro semi-automatico, basato sul Large Language Model (LLM), che assista il computista in questa associazione e confronti i risultati rispetto ad un insieme di metriche. Il flusso si articola in tre fasi. Nella prima vengono estratti dal file IFC gli oggetti e le relative proprietà; nella seconda viene fatta un'interrogazione del prezzario, organizzato in un database che individua le voci candidate per ciascun oggetto e il modello linguistico le ordina in base alle proprietà estratte; nella terza, dopo la conferma delle associazioni da parte dell'utente, quantità e costi vengono calcolati in modo deterministico. All’intelligenza artificiale (IA) non sono affidati né il calcolo delle quantità né la scrittura di codici e prezzi, che restano vincolati al modello informativo e al prezzario. La validazione è stata condotta sulla struttura di un edificio di quattro livelli, esportata in tre configurazioni IFC che differiscono per il contenuto informativo degli oggetti. I computi prodotti dal flusso sono stati confrontati con un computo manuale di riferimento redatto sul Prezzario Emilia-Romagna 2025, attraverso quattro famiglie di indicatori relative alla correttezza delle associazioni tra oggetti e voci, allo scarto sulle quantità, allo scarto sul costo e all'efficienza in termini di tempo e costo di elaborazione. Per ciascuna configurazione sono state eseguite quattro iterazioni, che corrispondono a versioni successive del flusso e non a ripetizioni indipendenti dello stesso test. Nell'ultima iterazione lo scarto sul costo totale rispetto al riferimento risulta inferiore all'1% in tutte e tre le configurazioni, con tempi di redazione compresi tra 76 e 134 minuti a fronte delle circa sei ore richieste dal computo manuale. Inoltre, il dato aggregato va letto insieme agli scarti sulle singole voci, più ampi e in parte compensati nel totale. Il flusso non restituisce una stima definitiva, ma un punto di partenza per il professionista, che viene alleggerito dei passaggi ripetitivi e il controllo preliminare, di norma a carico dell'operatore. La sperimentazione presenta alcuni limiti: il primo riguarda che il caso studio è un piccolo edificio della sola disciplina strutturale e il secondo è vincolato ad un abbonamento che permetta l’utilizzo del flusso di lavoro. Un possibile sviluppo è l'impiego di modelli eseguiti su un'infrastruttura locale, che garantirebbe la riservatezza dei dati di progetto e l'assenza di costi di abbonamento, con però una possibile riduzione della precisione dei risultati.
PROCEDURE AI DRIVEN DI ESTRAZIONE QUANTITÀ DA MODELLI IFC: PROBLEMATICHE DI VALIDAZIONE E DEFINIZIONE DI METRICHE VALUTATIVE
MUSSO, LEONARDO
2025/2026
Abstract
In the design of a public work, cost estimation is the phase on which the economic credibility of the project depends, since the bill of quantities and cost estimate (computo metrico estimativo) determines the base amount put out to tender. The Italian Public Procurement Code (Legislative Decree 36/2023) provides for the use of digital information management methods and tools for construction, as well as for the use of regional price lists to determine the amounts of the works items. Currently, widely used estimating tools do not allow the objects of the information model to be linked directly to price list items; the association remains largely manual, because it requires two semantic structures designed for different purposes to be made to communicate. The thesis aims to define and validate a semi-automatic workflow, based on a Large Language Model (LLM), that assists the quantity surveyor in this association, and to evaluate the results against a set of metrics. The workflow is structured in three phases. In the first, the objects and their properties are extracted from the IFC file; in the second, the price list, organised in a database, is queried to identify candidate items for each object, and the language model ranks them according to the extracted properties; in the third, after the user has confirmed the associations, quantities and costs are calculated deterministically. Artificial intelligence (AI) is entrusted neither with the calculation of quantities nor with the writing of codes and prices, which remain bound to the information model and to the price list. The validation was carried out on the structure of a four-level building, exported in three IFC configurations that differ in the information content of the objects. The estimates produced by the workflow were compared with a manual reference estimate prepared on the Emilia-Romagna 2025 Price List, through four families of indicators relating to the correctness of the associations between objects and items, the deviation in quantities, the deviation in cost, and efficiency in terms of processing time and cost. Four iterations were performed for each configuration; these correspond to successive versions of the workflow and not to independent repetitions of the same test. In the final iteration, the deviation in total cost from the reference is below 1% in all three configurations, with preparation times ranging from 76 to 134 minutes, compared with the approximately six hours required for the manual estimate. Moreover, the aggregate figure should be read together with the deviations on individual items, which are larger and partly offset each other in the total. The workflow does not deliver a definitive estimate but a starting point for the professional, who is relieved of repetitive steps and of the preliminary check normally carried out by the operator. The experimentation has some limitations: the first is that the case study is a small building covering the structural discipline only; the second is that the workflow depends on a subscription that enables its use. A possible development is the use of models run on local infrastructure, which would guarantee the confidentiality of project data and the absence of subscription costs, albeit with a possible reduction in the accuracy of the results.| File | Dimensione | Formato | |
|---|---|---|---|
|
Musso_Leonardo.pdf
Accesso riservato
Dimensione
4.82 MB
Formato
Adobe PDF
|
4.82 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/119057