Machine Learning Methods for Early CIPS Detection

The build-up of crud deposits on fuel in operating nuclear reactors has been a problem of concern for the past three decades that has resulted in fuel failures, power de-rate, and unplanned outages due to the phenomena of crud-induced-localized corrosion and crud-induced-power shift. Crud risk is currently managed using conservative fuel management and reactor operation strategies, resulting in higher fuel costs, due to the purchase of additional fuel, and higher maintenance costs due to periodic fuel cleaning. These costs are estimated to be $80 million per year across the national operating reactor fleet that can be significantly reduced. The efficient management of crud for the national reactor fleet is critical to the continued safe nuclear operation while eliminating a barrier for adopting new fuel products, such as accident tolerant fuel and high burnup fuel designed for longer fuel cycles and increased operating margin. This proposal will develop a methodology to predict and monitor the severity of crud for reactors, thereby allowing sufficient forewarning to address crud risk during the core reload design and subsequent operations. Monitoring of reactors for the onset of crud-induced power shift has the potential to significantly improve reactor operations and decrease fuel and operational costs. The methodology makes use of the continuous monitoring provided by the reactor fixed in-core detector system used by a significant number of reactors, a number that is growing as more plants move towards fixed detectors as part of their plant modernization programs. The method is based on a novel machine learning method that uses training datasets derived from a combination of measurements from fixed detectors and synthetic data. The synthetic data leverages the proven, high predictive accuracy of an existing software suite that simulates nuclear reactors. The proposed approaches the problem of fixed detector accuracy, currently limited to coarse data over 5-7 axial planes, by extending the measured data to fine-mesh (> 50 axial planes) that enables detection of local axial flux depressions associated with the onset of crud. The calculation of a novel crud index to assess crud risk based on in-core flux traces and determination of the crud index threshold value based on benchmark historical cycle of operation provides the integrated tool needed to not only monitor for crud impact but also design core loading patterns that minimize crud risk. In phase 1, the methodology for machine learning based on the use of generated synthetic data will be developed and demonstrated for the calculation of the crud index with application to core monitoring using fixed detectors. Phase 2 will focus on tightly integrating this methodology for an operating reactor into the existing crud screening recommended by industry experts. This would entail working closely with an industry partner to benchmark the reactor model using historical data and establish the crud index threshold based on cycles with and without crud-induced power shift.