Periodic Reporting for period 1 - iWeld (intelligent Weld inspection)
Reporting period: 2022-10-01 to 2024-03-31
The operational safety of nuclear power plants (NPPs) relies on maintaining the integrity of critical parts, especially those holding high pressure and operating at high temperatures. To withhold these conditions, the parts mentioned require joining with thick welds. During service, these components are subject to cyclic thermal and pressure loads, potentially leading to progressive damage such as fatigue cracks. Thick section welds play a similar role in other branches of industry, which use components operating under high pressure and high temperatures, such as oil and gas and off-shore. The potential damage scenarios and their inspection needs align with those in the nuclear context - both the loss of integrity and unnecessary repairs suggested by misinterpreted inspection incur high costs and significant disruption.
Ultrasonic methods have the greatest potential for inspecting welds because of their ability to penetrate the thick sections. Their basic principle is the use of the knowledge about the geometry and the wave velocity to realte the time of arrival to the location of sound reflectors (i.e. defects). Ultrasonic inspection commonly relies on the assumption of a homogeneous material structure with constant wave velocity, regardless of the direction of propagation and straight travel paths. However, this assumption is false for complex and heterogeneous materials. In anisotropic welds, the wave speed depends both on the local grain orientation and the propagation direction. Consequently, the ultrasound travels along curved paths, and the application of the homogeneity assumption in such welds leads to misinterpretations. For instance, an inspection may place the defect in the base material, while in reality, it is inside the weld or at its interface with the parent material. Moreover, the complex structure scatters the ultrasound, leading to reduced signal-to-noise ratio (SNR).
iWeld aims to address this issue through a process that can best be described as structure-informed imaging: By no longer assuming an isotropic and homogeneous structure, and taking the actual structure into account in the imaging process, it is possible to improve ultrasound imaging, both in terms of signal to noise ratio and in terms of defect localisation. In order to do this, we need to obtain reliable information about the local structure of the component under inspection. The traditional method is to dissect a mockup which has been manufactured using the same manufacturing procedure as the real component, and perform metallurgical analysis. This method is limited, as it usually uses simplified geometry, and is a single realization, unable to capture inevitable variations of influential parameters. iWeld therefore pursues two other avenues to obtain structure information : One approach called weld tomography using time of flight measurements across the weld to deduce structure information in an iterative approach, updating a discretized weld representation until calculated travel times converge with actual measurements. Obviously, a good starting point is essential for this to work.
A second approach uses weld simulation: Using data from the welding procedure, such as torch power, welding speed etc., it is possible to model the solidification and remelting during successive welding passes and to obtain a macroscopic weld structure description. This process is time consuming, and in order to capture the potential scope of outcomes due to inevitable variations of influential parameters, a large number of simulation would be required. To accelerate the process, the project proposes to use an AI-based meta-model, able to predict the actual weld structure from a limited number of actual weld simulations.
An important aspect of the project concerns the transfer and deployment of this approach to fields are than nuclear, which is vital to have a large panel of inspection service providers able to carry out structure informed ultrasound imaging. To that end, an advisory board with members from industries other than nuclear accompanies the project.
Ultrasonic methods have the greatest potential for inspecting welds because of their ability to penetrate the thick sections. Their basic principle is the use of the knowledge about the geometry and the wave velocity to realte the time of arrival to the location of sound reflectors (i.e. defects). Ultrasonic inspection commonly relies on the assumption of a homogeneous material structure with constant wave velocity, regardless of the direction of propagation and straight travel paths. However, this assumption is false for complex and heterogeneous materials. In anisotropic welds, the wave speed depends both on the local grain orientation and the propagation direction. Consequently, the ultrasound travels along curved paths, and the application of the homogeneity assumption in such welds leads to misinterpretations. For instance, an inspection may place the defect in the base material, while in reality, it is inside the weld or at its interface with the parent material. Moreover, the complex structure scatters the ultrasound, leading to reduced signal-to-noise ratio (SNR).
iWeld aims to address this issue through a process that can best be described as structure-informed imaging: By no longer assuming an isotropic and homogeneous structure, and taking the actual structure into account in the imaging process, it is possible to improve ultrasound imaging, both in terms of signal to noise ratio and in terms of defect localisation. In order to do this, we need to obtain reliable information about the local structure of the component under inspection. The traditional method is to dissect a mockup which has been manufactured using the same manufacturing procedure as the real component, and perform metallurgical analysis. This method is limited, as it usually uses simplified geometry, and is a single realization, unable to capture inevitable variations of influential parameters. iWeld therefore pursues two other avenues to obtain structure information : One approach called weld tomography using time of flight measurements across the weld to deduce structure information in an iterative approach, updating a discretized weld representation until calculated travel times converge with actual measurements. Obviously, a good starting point is essential for this to work.
A second approach uses weld simulation: Using data from the welding procedure, such as torch power, welding speed etc., it is possible to model the solidification and remelting during successive welding passes and to obtain a macroscopic weld structure description. This process is time consuming, and in order to capture the potential scope of outcomes due to inevitable variations of influential parameters, a large number of simulation would be required. To accelerate the process, the project proposes to use an AI-based meta-model, able to predict the actual weld structure from a limited number of actual weld simulations.
An important aspect of the project concerns the transfer and deployment of this approach to fields are than nuclear, which is vital to have a large panel of inspection service providers able to carry out structure informed ultrasound imaging. To that end, an advisory board with members from industries other than nuclear accompanies the project.
The first task in work package 1 consisted in compiling a list of industrially relevant thick section weld types. Each weld chamfer geometry is described by a parametric model, which is able to produce the input data for the welding simulation in work package 2.
Based on the matrix of relevant weld types, we have also produced a shortlist of actual welds representing different industry sectors, from which one weld was then selected for actual manufacturing. The specification of this demonstrator weld also contains a description of adequate instrumentation to be used during the manufacturing process, which will be used to validate and correct the welding simulation, if necessary. The mockup description was transferred to the workshop to instruct the actual manufacturing. At the end of the first reporting period, we were able to source the base and filler material and identify the automatic welding machine which will be used.
Workpackage 2 covers the welding simulations, and the development of an AI-based macrostructure predictor fed by these simulations. Based on the demonstration weld defined in WP1, we have developed a simulation approach using predefined weld beads. In this method, the weld bead contour is provided by the parametric weld model developed in WP1. The welding model was improved and is now able to handle up to 40 passes, which is sufficient for the demonstrator weld with a thickness of 32mm.
The results of these simulation will be used for the training of an AI model, which shall allow a more rapid prediction of weld macrostructures than the direct welding simulation can provide. The basic architecture of this model has been defined, using a continuously variable 3D mapping which describes the main anisotropic orientation within the weld. Furthermore, the interface between the welding simulation and the AI model was defined, which involves the condensation of the massive data volume produced by the welding simulation with a much lighter grain description.
Workpackage 3 covers the weld map tomography as an alternative path to obtain a weld structure description. We first carried out a number of finite element simulations to assess the effect of uncertainties in the weld description on the actual ultrasound imaging.
The actual weld map tomography model, which initially was only capable of pulse-echo ultrasound acquisitions with contact transducers, has been extended to tandem and immersion configurations. The tandem configuration turned out to be particularly challenging, as the ray paths cover only a limited portion of the weld, making the problem at hand quite ill-posed. Mitigation measures will be developed in the next reporting period.
Based on the matrix of relevant weld types, we have also produced a shortlist of actual welds representing different industry sectors, from which one weld was then selected for actual manufacturing. The specification of this demonstrator weld also contains a description of adequate instrumentation to be used during the manufacturing process, which will be used to validate and correct the welding simulation, if necessary. The mockup description was transferred to the workshop to instruct the actual manufacturing. At the end of the first reporting period, we were able to source the base and filler material and identify the automatic welding machine which will be used.
Workpackage 2 covers the welding simulations, and the development of an AI-based macrostructure predictor fed by these simulations. Based on the demonstration weld defined in WP1, we have developed a simulation approach using predefined weld beads. In this method, the weld bead contour is provided by the parametric weld model developed in WP1. The welding model was improved and is now able to handle up to 40 passes, which is sufficient for the demonstrator weld with a thickness of 32mm.
The results of these simulation will be used for the training of an AI model, which shall allow a more rapid prediction of weld macrostructures than the direct welding simulation can provide. The basic architecture of this model has been defined, using a continuously variable 3D mapping which describes the main anisotropic orientation within the weld. Furthermore, the interface between the welding simulation and the AI model was defined, which involves the condensation of the massive data volume produced by the welding simulation with a much lighter grain description.
Workpackage 3 covers the weld map tomography as an alternative path to obtain a weld structure description. We first carried out a number of finite element simulations to assess the effect of uncertainties in the weld description on the actual ultrasound imaging.
The actual weld map tomography model, which initially was only capable of pulse-echo ultrasound acquisitions with contact transducers, has been extended to tandem and immersion configurations. The tandem configuration turned out to be particularly challenging, as the ray paths cover only a limited portion of the weld, making the problem at hand quite ill-posed. Mitigation measures will be developed in the next reporting period.
An possible extension would be the inclusion of weld repairs, during which a section of the weld and/or base metal is grinded out and replaced by filler material. This changes the weld structure and chamfer geometry significantly.