Assessing the Trustworthiness of a Digital Twin within the Nearly Autonomous Management and Control System

Grubbs, Taylor; Shuppy, Addison; Athe, Paridhi; Son, Tran Cao; Dinh, Nam

Permalink

https://hdl.handle.net/2142/121794 Copy

Description

Title

Assessing the Trustworthiness of a Digital Twin within the Nearly Autonomous Management and Control System

Author(s)

• Grubbs, Taylor
• Shuppy, Addison
• Athe, Paridhi
• Son, Tran Cao
• Dinh, Nam

Issue Date

2023

Keyword(s)

• Autonomous control
• Digital twin
• AI trustworthiness
• Nuclear reactor

Abstract

Digital Twins have the capability to automate certain tasks and thereby improve the economics of advanced nuclear reactor operation and maintenance (O&M) activities. A key blocker to the acceptance of these technologies is the absence of a comprehensive framework for assessing the trustworthiness of digital twins. Due to the data-driven nature of Digital Twins, it is possible to harness the methods developed in the fields of Artificial Intelligence (AI) Safety to assess certain aspects of trustworthiness. However, this poses additional challenges. In the NAMAC system the Digital Twin for Diagnosis is trained from system-level thermal-hydraulic simulation data to predict a safety significant factor - peak fuel centerline temperature - during a Loss-of-flow scenario. Unfortunately, conventional AI safety techniques like Adversarial Testing are focused almost entirely on classification models based on static data - the Digital Twins within the NAMAC system are regression models based on time-series data. Symbolic rule extraction techniques for explainability suffers from the same classification model restriction- with the added difficulty of these techniques often being applicable only to certain model architectures. This study documents the first attempts at adapting these methods to the task of assessing the trustworthiness of the NAMAC digital twin for Diagnosis.

Type of Resource

text

Language

eng

Handle URL

https://hdl.handle.net/2142/121794

Owning Collections