How a digital twin for process safety will differ from a production twin

Journal Article (2026)
Author(s)

Kamran Gholamizadeh (Risk Analytics AI , University of Quebec)

Stewart W. Behie (Texas A&M University Corpus Christi)

Hans J. Pasman (TU Delft - Old - sect Particle Technology (DCT/PART))

Ming Yang (TU Delft - Technology, Policy and Management)

Research Group
Safety and Security Science
DOI related publication
https://doi.org/10.1016/j.psep.2026.109780 Final published version
More Info
expand_more
Publication Year
2026
Language
English
Research Group
Safety and Security Science
Journal title
Process Safety and Environmental Protection
Volume number
222
Article number
109780
Page Views
2
Reuse Rights

Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.

Abstract

To alleviate complexity, digital twins (DTs) have moved rapidly from a concept to a widely promoted manufacturing support technology across various types of industry, including the process industries. The literature offers abundant demonstrations of DTs for manufacturing used to monitor the functioning of the process assets to provide information on anomalies. So, it will be great for production optimization and training operators to control upsets. With this, safety is at stake, yet the conditions under which DTs can be trusted to inform safety-critical decisions remain poorly defined. No structured account is available of what distinguishes a safety-qualified digital twin from a sophisticated process safety control dashboard. This paper asserts that such a safety twin addressing that gap requires five functional types of twins for their primary safety role. Further, validation evidence is required before deployment and autonomy may be granted to them, and the governance obligations they carry. The typology is mapped onto seven elements of process safety management (PSM). For its realization, a six-layer reference architecture is proposed that spans physical data acquisition, process modelling and plant degradation, predictive analytics and fault diagnosis, dynamic risk assessment and safety intelligence, human-centred decision support, and governance, together with a tiered validation hierarchy anchored in functional safety practice. Finally, two reproducible Python case demonstrations ground the framework: 1) a multivariate anomaly-detection for a simulated continuous distillation unit, which detected an abrupt condenser fault within 8 min and two developing faults within operationally useful horizons at a false-alarm rate of one event per day, and 2) a digitised Bow-Tie for a pressurised hydrocarbon reactor, in which barrier health degradation raised the top-event likelihood proxy by 21.2% over 55 months and in which a sensitivity analysis isolated the single most risk-critical barrier. The paper closes with a critical appraisal of data dependency, uncertainty communication, and the organisational gap between demonstration and deployment.