This work has demonstrated the use of the QClimax machine learning package for fitting of QENS experimental data.
Significance and Impact
The improvement of QClimax's machine learning global fits over classical models has been demonstrated via application to experimental data.
- Fit ionic fluid experimental data with physical constraints on the model.
- Compared the result to naive, unconstrained, local fittings to demonstrate the increased accuracy.
Modern Quasi-Elastic Neutron Scattering(QENS) spectrometers produce experimental data at high energy ranges that cannot be accurately described by a single component. Rather, the data is binned into discrete Q values, which are then fit by multiple components defined by formulas with a small number of variables. Classical models fit these bins on their own, which loses the dependence of the component parameters on Q. The QClimax software package, part of the Integrated Computational Environment for Modeling and Analysis of Neutrons (ICEMAN) workbench, avoids this by fitting all bins at once and allowing a functional constraint to be defined on parameters.
QClimax has been used in the creation of a model for the room temperature ionic liquid H2NC(dma)2 for comparison to the classical method of sequential fitting. In both cases, two Lorentzian components were used alongside a linear background component. A jump diffusion model is then imposed onto the Lorentzians' parameters. As demonstrated in the plots to the right, the imposition of the jump diffusion constraint within the global fitting produced a model that, while slightly less good at some individual Q values, was ultimately much better fit to the experimental data than the original jump diffusion, while avoiding the over-fitting caused by ignoring the parameter Q dependence, which cause physically incorrect results for the model in the sequential, unconstrained case.