Application of a Machine Learning Approach Based on Nearest Neighbors to Extract Embrittlement Trends from RPV Surveillance Data

Abstract

Over the past 10-15 years considerable effort by many different researchers has produced codified embrittlement trend curves for use in calculations that support the safe life management of nuclear reactor pressure vessels. This paper explores the complementary insights made possible by a more localized approach to data fitting using nearest-neighbor techniques. Our objective is to evaluate the accuracy and precision of nearest neighbor predictions of transition temperature shift (ΔT41J) using a comprehensive database of international surveillance data. In the nearest neighbor method, a distance vector is used to quantify the proximity (and thereby the similarity) of all known conditions to some condition of interest. This paper evaluated the effect of different distance vector and partition criteria choices. Higher dimensional distances appear to offer the promise of greater embrittlement trend fitting accuracy. However, this accuracy reduces the ability to apply the method to all conditions encountered in the fleet. As our goal is to identify data from one plant that can be used to inform embrittlement trends in other plants (e.g., as is done in integrated surveillance programs), it seems appropriate to adopt the simplest distance vector that offers reasonable accuracy (i.e., copper, nickel, temperature). We combined this distance, which ranks the similarity of irradiation sensitivity between different conditions, with a copper measurement precision of ± 0.03 mass% that serves as a trigger to determine the (Cu, Ni, T) distance to the partition limit (called ρ). When used with the surveillance database, this approach indirectly limits the nickel and temperature variation within the partition, making explicit limits on nickel and temperature unnecessary. While some effect of copper on ΔT41J within the partition remains, it has little practical effect on the mean nearest neighbor trend curve while slightly widening the tolerance bounds, which can be viewed as a conservatism.