Testing Generative Pre-trained Transformer for Knowledge Abstraction and Reasoning for Nuclear Reactor Design and Safety Applications

Athe, Paridhi; Dinh, Nam

Permalink

https://hdl.handle.net/2142/121795 Copy

Description

Title

Testing Generative Pre-trained Transformer for Knowledge Abstraction and Reasoning for Nuclear Reactor Design and Safety Applications

Author(s)

• Athe, Paridhi
• Dinh, Nam

Contributor(s)

U.S. Department of Energy

Issue Date

2023

Keyword(s)

• Generative pre-trained transformers
• Knowledge abstraction and reasoning
• Nuclear reactor

Abstract

The decision regarding the adequacy of a Modeling and Simulation (M & S) tool is governed by different challenging factors. The M & S tools significantly vary in fidelity and scale of the simulation. Moreover, representative data for validating the adequacy of the code is scarce and limited. Therefore, standard procedures and methodologies, like, Code Scaling Applicability and Uncertainty (CSAU) and Evaluation Model Development and Assessment Process (EMDAP) have been developed by the United States Nuclear Regulatory Commission (US NRC) to guide the development and assessment of modeling and simulation tools. Though comprehensive in scope, different elements of these methodologies are governed by expert input and insight. Phenomena Identification and Ranking Table (PIRT), in particular, is based on expert input/opinion regarding the importance and knowledge level of different phenomena with respect to the figures of merit. The quality of PIRT depends on the domain knowledge and expertise of the participating group of experts. Knowledge abstraction and reasoning are important for PIRT formulation and understanding relationships and dependencies between different phenomena. The decision of adequacy of an M & S tool for an intended use also requires information retrieval, context-based knowledge abstraction, and reasoning to provide adequate evidential support. The risk assessment for nuclear reactor applications is based on probabilistic analysis and requires an estimation of the safety margin while considering different heterogeneous factors related to verification, validation, and uncertainty quantification of M & S tools. This process of knowledge abstraction and reasoning can be quite time-consuming, costly, and resource incentive. In this work, we explore and test the capabilities of Artificial intelligence-based large language models called Generative Pre-trained Transformers in information retrieval, knowledge abstraction, and reasoning to support evidence-based adequacy analysis of M & S tools. In particular, we analyze the ability and efficiency of GPT in replacing the expert-driven elements in adequacy analysis of M & S tools for an intended application (e.g., formulation of PIRT, understanding relationship and interdependencies between different phenomena. The main objective of this paper is to test the domain knowledge of GPT and identify its strengths, weaknesses, and challenges for their application in the nuclear engineering domain.

Type of Resource

text

Language

eng

Handle URL

https://hdl.handle.net/2142/121795

Sponsor(s)/Grant Number(s)

Federal Research and Development Funding

Owning Collections