Non-judgemental Dynamic Fuel Cycle Benchmarking

TL;DR

This paper introduces a novel benchmarking methodology combining Gaussian process regression and dynamic time warping to compare fuel cycle simulations, accounting for uncertainty and avoiding common pitfalls in signal processing.

Contribution

It develops a new approach for fuel cycle benchmarking that integrates machine learning and time series analysis, providing robust figures-of-merit without requiring common time grids.

Findings

The methodology effectively compares simulator performance across metrics.

It ranks metric importance using a derived contribution measure.

Gaussian process regression handles error reduction better than standard techniques.

Abstract

This paper presents a new fuel cycle benchmarking analysis methodology by coupling Gaussian process regression, a popular technique in Machine Learning, to dynamic time warping, a mechanism widely used in speech recognition. Together they generate figures-of-merit that are applicable to any time series metric that a benchmark may study. The figures-of-merit account for uncertainty in the metric itself, utilize information across the whole time domain, and do not require that the simulators use a common time grid. Here, a distance measure is defined that can be used to compare the performance of each simulator for a given metric. Additionally, a contribution measure is derived from the distance measure that can be used to rank order the importance of fuel cycle metrics. Lastly, this paper warns against using standard signal processing techniques for error reduction. This is because it is found that error reduction is better handled by the Gaussian process regression itself.