An NLP-driven analytical pipeline for exploring Digital Twin research maturity in construction

Purpose Conventional bibliometric techniques for analysing Digital Twin (DT) research maturity in the architecture, engineering, and construction (AEC) industry rely on keyword frequency and overlook semantic relationships and system-level technology interactions. This study develops an integrated analytical pipeline for measuring DT research maturity through semantic modelling, explanatory analysis and configuration search. Design/methodology/approach This study developed a multi-stage analytical pipeline integrating Sentence-BERT (SBERT)-based semantic filtering, BERTopic topic modelling and maturity classification, extended by clustering, SHapley Additive exPlanations (SHAP)-based explanatory modelling and heuristic optimisation. Using a corpus of 791 semantically filtered publications, the framework generates a Digital Twin Maturity Index (DTMI). Each analytical stage is reported with its full specification, and the index, the attribution stage and the optimisation stage are examined for dependence on their construction, including a seed and specification sensitivity analysis of the optimisation over 200 runs. Findings The corpus shows a shift from broadly framed early work towards technology-specific research, with sensors, building information modelling (BIM), artificial intelligence (AI) and Internet of Things the dominant tags and Cybersecurity, 5G and Blockchain the least consolidated. The three-component DTMI rises from −2.07 in 2017 to 0.73 in 2026, but publication year is one of its three components, so that rise is partly an arithmetic property of the index rather than evidence of maturation. The SHAP stage identifies sensors, AI and BIM as the strongest correlates of DTMI, an ordering that tracks corpus prevalence because the index contains the technology count as a component. The heuristic optimisation stage does not survive sensitivity analysis: 99.6% of candidate states are accepted, the final state is essentially uncorrelated with the data-derived starting point and across 200 seeds every technology contributes between 6.4% and 8.9% on average, so the reported prominence of geographic information system (GIS) and Cloud reflects a single draw rather than a recovered pattern. Originality/value This study contributes a replicable NLP- and ML-driven framework integrating semantic discovery, maturity measurement and explanatory analysis within a single pipeline, reported at a level of specification that allows each stage to be independently checked. Its second contribution is diagnostic: it shows that a composite maturity index containing a temporal term will trend by construction, that attribution against such an index recovers corpus prevalence rather than maturity drivers and that a stochastic search over a hand-specified objective can produce a confident technology ranking that is indistinguishable from noise. These failure modes are common in the digital-technology literature and are rarely tested for. The findings are literature-derived signals for framing research agendas; the study analyses publications rather than project or firm outcomes and provides no basis for investment or policy guidance.