Periodic Reporting for period 1 - ASSAS (Artificial intelligence for the Simulation of Severe AccidentS)
Période du rapport: 2022-11-01 au 2024-04-30
The ASSAS (Artificial intelligence for the Simulation of Severe AccidentS) project aims to foster the development of accurate and computationally efficient nuclear power plant simulators with severe accident capabilities.
Different types of simulators are used in the nuclear sector for education, engineering, safety analysis, and operator training. The most realistic simulators replicate the main control room of a given reactor and reproduce the response of the plant in real time. Other simulators display a graphical user interface (GUI) to interact with the simulation from a computer screen. However, most of them do not have severe accident modelling capabilities, which means that cannot simulate situations with extended fuel degradation, like in the Fukushima-Daiichi or Three Miles Island accidents. ASSAS plans to address this shortcoming.
Severe accident calculation codes like ASTEC (Accident Source Term Evaluation Code, developed by IRSN) and Melcor (developed by Sandia National Laboratory for the Nuclear Regulatory Commission of the USA) can model the global plant behaviour during the different phases of severe accidents. However, they are rarely interfaced with simulators, which would have several advantages for severe accident experts. First, it would allow users to rely on a more detailed description of the plant systems. In addition, having a graphical user interface would improve the learning curve for new severe accident code users, which is a burning issue because of the renewal of the workforce in this field. In the frame of ASSAS, ASTEC will be interfaced with simulation platform of Westinghouse Spain named TEAM_SUITE®, to form a basic-principles severe accident simulator. This desktop simulator will feature a 4-loop Western-type Pressurized Water Reactor (PWR). It will include two severe accident sequences to exercise the user:
- a large-break loss-of-coolant accident with safety injection and containment spray failure,
- a station black-out with auxiliary feedwater failure.
This proof-of-concept will shows the possibility to develop more detailed simulators with extended capabilities and with other reactor designs.
The main scientific challenge for the development of severe accident simulators is to enable their execution in real-time (and if possible several times faster), to give a realistic experience to the user. Most research performed in ASSAS will tackle this issue, while keeping an acceptable accuracy of the calculations. Different approaches will be explored:
- The simplification of the physical models,
- The optimisation of ASTEC’s algorithmic, especially its parallelisation,
- The use of machine-learning.
Large resources will be dedicated to this last strategy, which is the most innovative one. Partners will team up to generate a training database composed of precalculated simulations, from which the machine-learning models will learn to reproduce the results generated by physical models. Once trained, such models are several orders of magnitude faster than their physical counterparts. Different approaches will be evaluated to replace some calculation steps (with ASTEC only) or the severe accident codes completely (with ASTEC and Melcor). The former methodologies, called hybrid approaches, are only possible with ASTEC because the consortium has access to its source code. Severe accident phenomenology is highly complex, with strong couplings between various physical phenomena and extreme non-linearities. Using artificial intelligence methods so extensively for severe accident modelling has never been done before.
The feedback from the project will be of high value for other multi-physics simulation codes, in the nuclear industry and beyond. The training database generated during the project will be openly accessible to the scientific community.
Different types of simulators are used in the nuclear sector for education, engineering, safety analysis, and operator training. The most realistic simulators replicate the main control room of a given reactor and reproduce the response of the plant in real time. Other simulators display a graphical user interface (GUI) to interact with the simulation from a computer screen. However, most of them do not have severe accident modelling capabilities, which means that cannot simulate situations with extended fuel degradation, like in the Fukushima-Daiichi or Three Miles Island accidents. ASSAS plans to address this shortcoming.
Severe accident calculation codes like ASTEC (Accident Source Term Evaluation Code, developed by IRSN) and Melcor (developed by Sandia National Laboratory for the Nuclear Regulatory Commission of the USA) can model the global plant behaviour during the different phases of severe accidents. However, they are rarely interfaced with simulators, which would have several advantages for severe accident experts. First, it would allow users to rely on a more detailed description of the plant systems. In addition, having a graphical user interface would improve the learning curve for new severe accident code users, which is a burning issue because of the renewal of the workforce in this field. In the frame of ASSAS, ASTEC will be interfaced with simulation platform of Westinghouse Spain named TEAM_SUITE®, to form a basic-principles severe accident simulator. This desktop simulator will feature a 4-loop Western-type Pressurized Water Reactor (PWR). It will include two severe accident sequences to exercise the user:
- a large-break loss-of-coolant accident with safety injection and containment spray failure,
- a station black-out with auxiliary feedwater failure.
This proof-of-concept will shows the possibility to develop more detailed simulators with extended capabilities and with other reactor designs.
The main scientific challenge for the development of severe accident simulators is to enable their execution in real-time (and if possible several times faster), to give a realistic experience to the user. Most research performed in ASSAS will tackle this issue, while keeping an acceptable accuracy of the calculations. Different approaches will be explored:
- The simplification of the physical models,
- The optimisation of ASTEC’s algorithmic, especially its parallelisation,
- The use of machine-learning.
Large resources will be dedicated to this last strategy, which is the most innovative one. Partners will team up to generate a training database composed of precalculated simulations, from which the machine-learning models will learn to reproduce the results generated by physical models. Once trained, such models are several orders of magnitude faster than their physical counterparts. Different approaches will be evaluated to replace some calculation steps (with ASTEC only) or the severe accident codes completely (with ASTEC and Melcor). The former methodologies, called hybrid approaches, are only possible with ASTEC because the consortium has access to its source code. Severe accident phenomenology is highly complex, with strong couplings between various physical phenomena and extreme non-linearities. Using artificial intelligence methods so extensively for severe accident modelling has never been done before.
The feedback from the project will be of high value for other multi-physics simulation codes, in the nuclear industry and beyond. The training database generated during the project will be openly accessible to the scientific community.
The specifications of the basic-principles simulator developed in ASSAS have been defined collegially by consortium members, with the help of the end-user group and the advisory board. The sequences have been chosen to show the main phenomena involved in severe accidents. The most important information to be displayed to the user to show the progression of the accident has been selected.
The corresponding reactor model has been developed for ASTEC. It has been optimised to feature the best compromise between accuracy and rapidity. Combined with algorithmic improvements, this work has led to a three-fold reduction in execution time.
In parallel, a solution has been developed to interface ASTEC with Westinghouse Spain simulation platform, TEAM_SUITE®, which has already been successfully tested with a simplified model.
Important steps have been achieved to prepare the development of machine-learning models. The architecture of the database hosting the training data has been defined and partially developed. The different strategies to be explored during the project have been described in detail, to give clear specifications to the model developers. Some tests have been carried out with sample data to explore the potential of various approaches. When necessary, ASTEC’s source code has been modified to make the connection with machine-learning models possible.
The corresponding reactor model has been developed for ASTEC. It has been optimised to feature the best compromise between accuracy and rapidity. Combined with algorithmic improvements, this work has led to a three-fold reduction in execution time.
In parallel, a solution has been developed to interface ASTEC with Westinghouse Spain simulation platform, TEAM_SUITE®, which has already been successfully tested with a simplified model.
Important steps have been achieved to prepare the development of machine-learning models. The architecture of the database hosting the training data has been defined and partially developed. The different strategies to be explored during the project have been described in detail, to give clear specifications to the model developers. Some tests have been carried out with sample data to explore the potential of various approaches. When necessary, ASTEC’s source code has been modified to make the connection with machine-learning models possible.
ASSAS will have various results. The basic-principles simulator will be the most tangible deliverable of the project. The objective is to have a prototype that can be used for educational purposes about severe accidents, showing the phenomenology in an interactive manner. This proof of concept will also give confidence to industrials to extend their existing engineering and full-scale simulators with severe accident capabilities.
Others results are also of high significance. An optimised version of ASTEC with improved performances will be delivered to all users, potentially with highly performant machine-learning models. This will not only enhance the productivity of severe accident research and engineering, but potentially also open the way to new approaches and tools that are today out of reach. The return from experience concerning these models can be exploited beyond the nuclear community: all fields of engineering using multi-physics codes may be concerned.
The database used to train artificial intelligence models will be made openly accessible, for other researchers to test new approaches on large and high-quality data.
Others results are also of high significance. An optimised version of ASTEC with improved performances will be delivered to all users, potentially with highly performant machine-learning models. This will not only enhance the productivity of severe accident research and engineering, but potentially also open the way to new approaches and tools that are today out of reach. The return from experience concerning these models can be exploited beyond the nuclear community: all fields of engineering using multi-physics codes may be concerned.
The database used to train artificial intelligence models will be made openly accessible, for other researchers to test new approaches on large and high-quality data.