报告人：Svetlana Mikhalycheva 研究员 (白俄罗斯国家科学院斯捷潘诺夫物理研究所)

时 间：2024年9月18号（周三）14：00-16：00

地 点：线下-良乡校区，物理实验中心229会议室

摘要：

Computer-based automated data analysis procedures are not only much faster and less demanding for expert knowledge of operators than manual techniques, but also provide reliability and reproducibility of the analysis results due to much lesser vulnerability to accidental errors. Generally, analysis can be based on the previously gained experience (human expertise for manual approaches or machine learning for computer systems) or rely on fundamental principles of physics, mathematics, statistics, etc., when the analysis algorithms are derived instead of being “guessing”. In this lecture, we consider the latter approach: namely, we demonstrate the capabilities of two efficient mathematical tools (Bayes’ formula and Fisher information) suitable for automatic experiments planning and data analysis imaging, X-ray diffractometry, and mass-spectrometry.

Fisher information quantifies the ability to reconstruct model parameters from measured data with a statistical noise, giving the information about the reliability of the obtained results and the possible correlations between the model features. We show that this tool is suitable for optimizing the planned high-resolution X-ray diffraction measurements by choosing the most informative reflections and measurement geometries depending on the sample parameters of interest. Also, banded diagonal structure of the inverse Fisher matrix implies that the reconstruction problem can be solved locally, by iterative consideration of only a subset of parameters to be reconstructed at each step. We apply this technique to the problem of decomposing an X-ray and mass spectra, consisting of a large number of strongly overlapping peaks, into contributions of individual peaks. In application to super-resolution optical fluctuation imaging (SOFI), minimization of the predicted reconstructed error according to Fisher information helps us to choose the optimal sample-dependent order of cumulant image. To improve the performance of the approach, we also demonstrate application of artificial neural network, trained at the information-based results, to prediction of the optimal cumulant order.

Bayes’ theorem quantitatively connects our a posteriori knowledge about the investigated sample with a priori knowledge and the measurement results and is applicable to ranking of models in the pattern recognition problem. We apply a Bayesian approach to quantitative analysis of X-ray powder diffraction data and mass spectra (identifying contributions from separate phases or ions in the sample response), and to peak indexing in high-resolution X-ray diffraction profiles (finding the correspondence between each measured peak with an expected Bragg peak, a layer thickness oscillation fringe or a superlattice fringe).

简历：

Svetlana Mikhalycheva is a senior researcher in B.I.Stepanov Institute of Physics，National academy of sciences of Belarus. She has published more than 20 papers in top journals in physics, such as Phys.Rev , npj Quantum Inf. and so on. Her research covers: classic simulation of quantum state. She has finished projects like EU Horizon 2020 project, BRFFR-KOREA.

联系方式：shenglizhangustc@126.com

邀请人： 张胜利

网 址：/

承办单位：物理学院、先进光电量子结构设计与测量教育部重点实验室