16th International Conference on Muon Spin Rotation, Relaxation and Resonance (μSR2025)

μSR Data Analysis with Bayesian Neural Networks

Not scheduled

20m

Poster Presentation Beamlines and instruments Poster Session 1

Speaker

Mr Miyahara Hiroaki (Muroran Institute of Technology)

Description

In muon spin rotation/relaxation ($\mu$SR) experiment, the time evolution of muon spin polarization, $P(t)$, is analyzed to evaluate the internal magnetic fields and their temporal fluctuations sensed by the muons at their stopping sites. In conventional data analysis, a phenomenological model for $P(t)$—typically based on general spin relaxation theory (e.g. that by Kubo and Toyabe) —is used in the least-squares fits to deduce parameters for reproducing $P(t)$. These parameters are then interpreted in terms of the physical models specific to the materials. This approach often requires trial and error in selecting the phenomenological model and establishing its correspondence with the physical model, demanding significant experience and time before yielding results. Such an empirical approach, heavily relying on human intervention, is expected to become increasingly impractical in the face of unprecedented volumes of data generated by the development of new data acquisition methods that leverage the high beam intensity from J-PARC's 1 MW-class operation [1]. Therefore, there is a growing need for fast and reliable analysis methods that take advantage of rapidly advancing artificial intelligence (AI).
We are attempting to apply AI-based regression analysis techniques to $\mu$SR data analysis. Specifically, our approach is based on deep learning using neural networks (NNs), where the challenge has been to quantitatively evaluate the uncertainty of predicted parameters in phenomenological models. We have tackled this issue by using Bayesian Neural Networks (BNNs), which incorporate Bayesian inference into NNs [2]. We found that the distribution in NNs predictions evaluated by BNNs serves as a useful measure for the prediction errors.
In this presentation, we report the results of applying Monte Carlo Dropout [3] to NNs to approximate BNNs, where μSR data observed in a hybrid-organic-inorganic-perovskite HC(NH₂)₃PbI₃ are analyzed by an AI which has been trained by the extended Kubo-Toyabe function [4].

References:
[1] S. Nishimura, et al., Nucl. Instrum. & Meth. Phys. Res. A 1056, 168669 (2023).
[2] R. M. Neal, Springer Lecture Notes in Statistics 118, 1 (1996).
[3] Y. Gal and Z. Ghahramani, Proc. of The 33rd Inter. Conf. on Mach. Lear.: Proc. Mach. Lear. Rese. (PMLR) 48, 1050-1059 (2016).
[4] T. U. Ito and R. Kadono, J Phys. Soc.Jpn 93, 044602 (2024).