Automated formalization of expert knowledge for data-knowledge-fused digital twin model extraction using pre-trained language models

Integrating expert knowledge into Digital Twin model extraction remains an open challenge. Although expert knowledge can enhance model robustness and contextual validity, it is typically expressed in unstructured natural language, leading to ambiguity, heterogeneity, and limited operationalizability. Consequently, systematic and automated formalization mechanisms are required to enable effective fusion of expert knowledge with data for extracting data-knowledge-fused Digital Twin models. To address this gap, we propose an automated approach for formalizing expert knowledge expressed in natural language. We evaluate two formalization approaches using twelve publicly available pre-trained language models under a set of explicitly defined constraints. Based on the empirical evaluation, encoder-decoder language models emerge as the most suitable candidates, as they demonstrate the strongest overall performance within the considered evaluation setting. We further illustrate the role of the proposed approach within data-knowledge-fused Digital Twin model extraction through a reliability case study. The proposed approach reduces effort required for integrating expert knowledge and supports automated model extraction.