The F-t-Pj-RG method: An adjacent-rolling-windows based steady-state detection technique for application to kinetic Monte Carlo simulations

• Published in: Computer physics communications 2018-11, Vol.232 (C), p.124-138
• Main Authors: Nellis, Chris, Danielson, Thomas, Savara, Aditya, Hin, Celine
• Format: Article
• Language: English
• Subjects: Complex chemical reactions; Kinetic Monte Carlo simulations; MATHEMATICS AND COMPUTING; Steady-state detection
• ISSN: 0010-4655; 1879-2944
• DOI: 10.1016/j.cpc.2018.05.013

A window-based steady-state detection algorithm has been developed for application to kinetic Monte Carlo simulation data. The algorithm, termed F-t-Pj-RG sequentially applies an F-test, a t-test, and a projection test on adjacent windows of the data while rolling (or shifting) and growing the windows when any of the tests fail. In aggregate, the algorithm is able to (a) automatically reject the warm-up period as not being at steady-state, as well as (b) determine an appropriate window size for converged statistics when sampling the data, which is necessary for detection of steady-state, and (c) detect steady-state within a particular tolerance. The last step, the projection test, is actually an oscillating-slope projection test, and is performed on j sequential data windows (i.e., more than two adjacent windows). It requires more than simply being within the user defined tolerance: the oscillating-slope projection test includes a condition that the slope must oscillate around zero when ≥2, which is an additional indication of steady-state. When all three tests are passed, the F-t-Pj test is passed, indicating that the prerequisites of steady-state detection have been met and also that conditions consistent with the definition of steady-state have been realized. This algorithm is applied to a variety of data sets that correspond to the diverse type of data trends that can be produced by kinetic Monte Carlo simulations. The algorithm is shown to be robust in its ability to handle differing functional forms, and is able to detect steady-state with low computational cost. The low computational cost of this method and its robustness towards varied data trends make it suitable for on-the-fly use in kinetic Monte Carlo simulations.