Combining AI and Physics-Based Digital Twins in Nuclear Industries

Albati, Mohammad; Bui, Ha; Sakurahara, Tatsuya; Mohaghegh, Zahra

Permalink

https://hdl.handle.net/2142/121842 Copy

Description

Title

Combining AI and Physics-Based Digital Twins in Nuclear Industries

Author(s)

• Albati, Mohammad
• Bui, Ha
• Sakurahara, Tatsuya
• Mohaghegh, Zahra

Issue Date

2023

Keyword(s)

• Digital twins
• machine learning
• simulation
• artificial intelligence

Abstract

Many next-generation advanced nuclear reactors will likely use Artificial Intelligence (AI) to enhance safety and reduce the cost of operation and maintenance of the nuclear power plants (NPPs). As an example, Digital Twins (DTs) can utilize AI/ML technology for verifying simulations and analyzing data trends to recommend decisions or perform actions. The development of DT models requires an understanding of the system (such as governing equations and system purpose, etc.) that it replicates in order to simulate the behavior of that system correctly. There are two approaches in which DT models can be built. First is the Data-Driven approach. This approach uses machine learning algorithms such as neural networks to predict the behavior of the system (represented by a specific Key Performance Metric KPM such as outlet coolant temperature) based on the physical features of the system such coolant flow, inlet temperature and reactor power. To train the machine learning model, vectors of input features (such as coolant flow, inlet temperature and reactor power) and output (independent variable) such as outlet coolant temperature are needed. The data to train the model can be obtained from real world observations (using sensors) or using high fidelity physics -based simulations. Irrelevant to the source of the training data, this approach is called data-driven approach. The second approach to build the digital twin is to use simulations (high fidelity simulations or surrogate models) to predict the behavior of the system. Each approach has its own drawbacks. There are many situations and physical processes occurring in NPPs that do not lend themselves to first principles modeling (e.g., fuel performance, piping erosion/corrosion), because the process is too complex to run a first principles model, or the underlying physical mechanism is not understood well enough. Data-driven approaches can solve this problem using a machine learning algorithm that learns from real world observations to build the model. However, the validity of the model is only limited to the range of the training data. Simulation-based approaches can predict more variables that cannot be measured using sensors such as the heat inside the reactor core. However, simulations require substantial understanding of the physical phenomena and simulation tools may require extensive computational resources. The combination of both data-driven and physics-based models in one application can be beneficial to cover up for the drawbacks of using each system separately. The aim of this paper is to: i) conduct a literature review on the use of DT in nuclear industry, ii) propose a methodology of utilizing a mixed approach (data-driven plus surrogate models) to build the DT model, iii) use an AI algorithm to build the data-driven portion of the DT model, and iv) apply the methodology to a hypothetical system as demonstration of concept.

Type of Resource

text

Language

eng

Handle URL

https://hdl.handle.net/2142/121842

Owning Collections