Display Accessibility Tools

Accessibility Tools

Grayscale

Highlight Links

Change Contrast

Increase Text Size

Increase Letter Spacing

Readability Bar

Dyslexia Friendly Font

Increase Cursor Size

Comprehensive Exam of CMSE Zhiyuan She

Department of Computational Mathematics, Science & Engineering

Michigan State University

Comprehensive Exam Notice

 

Sep. 27th 1:00pm - 2:30 pm, 1420EB

Zoom link: https://msu.zoom.us/j/2338589328

Passcode: 239733

 

Machine Learning-based Stochastic Reduced Modeling of General Langevin Equations (GLE) and State-dependent-GLE

Zhiyuan She

 


Abstract:

An important challenge in constructing the reduced dynamics of molecular systems is accurately modeling the non-Markovian behavior that arises from the dynamics of unresolved variables. This issue is primarily due to the lack of scale separation, leading to reduced dynamics that typically exhibit significant memory effects and non-white noise terms. We propose two data-driven approaches to learn reduced models for multi-dimensional resolved variables while faithfully capturing non-Markovian dynamics. Unlike common methods that directly construct the memory function, our approach identifies a set of non-Markovian features that encapsulate the history of the resolved variables. We then establish a joint learning framework for the extended Markovian dynamics involving both the resolved variables and these features. The training process involves matching the evolution of the correlation functions of the extended variables, which can be directly derived from the correlation functions of the resolved variables. The first model we develop approximates the multi-dimensional generalized Langevin equation and ensures numerical stability without requiring empirical adjustments. We validate the effectiveness of this method by applying it to construct reduced models for molecular systems with one-dimensional and four-dimensional resolved variables.  The second model considers the dependence of the memory kernel and mass on position. We demonstrate that this second approach outperforms the first model through examples involving one-dimensional and two-dimensional resolved variables. Future work will connect our current work with stochastic control problems.

 

 

Committee:

Huan Lei (chair)

Di Liu

Michael Murillo

Yimin Xiao