题目:Machine Learning for Scientific Computing
主讲人:Akeel Shah 教授 重庆大学 能源与动力工程学院
邀请人:邢炜博士 酷游ku真人 微电子学院
Abstract
Machine learning surrogate models are widely used for applications such as design optimization and uncertainty quantification, where repeated evaluations of an expensive simulator are required. In many real applications, the model inputs and/or outputs are spatial or spatiotemporal (possible stochastic) fields and thus a direct implementation of the standard surrogate model is often unfeasible. It is a common practice to find approximate representations for the inputs and outputs based upon linear dimension reduction methods or simplified covariance structures. The challenges are 1) to build models that are capable of very high dimensional input and/or output learning; 2) to efficiently capture the complex nonlinear correlations across spatial dimensions and input parameters; and 3) multi task learning or missing/noisy data. In this talk I will discuss a number of approaches, based on manifold learning, proper orthogonal decomposition, Karhunen Loeve expansions, and tensorial representations with shared input-output latent spaces. I will discuss multi fidelity approaches based on residual stochastic process or deep learning models. Ab initio, microscopic, mesoscale and multiscale modelling, which also suffer from high computational costs, is a natural extension. Initial ideas in this regard will be discussed briefly.
Short Biography
Professor Akeel Shah graduated with a first-class honours degree in Mathematical Physics in 1995 and a PhD in Applied Mathematics (both from University of Manchester Institute of Science and Technology) in 2001. He is currently a Professor in the School of Energy and Power Engineering at the University of Chongqing, with expertise in electrochemical energy conversion, computational engineering and applied machine learning. He previously held positions at University of Southampton and University of Warwick. His work is primarily focused on the modelling and simulation of energy-conversion devices (flow batteries, metal-air batteries, organic/inorganic fuel cells), including computational modelling, and the development of fast algorithms for computer codes in science and engineering based on machine learning and computational statistics. Between 2004 and 2006, he held a joint Pacific Institute of Mathematics Sciences (PIMS) and Mathematics of Information Technology and Complex Systems (MITACS) Fellowship. He is the author of over 75 publications in leading, international peer-reviewed journals. Professor Shah has worked closely with the fuel cell and battery industry (Ballard Power Systems, Johnson Matthey Plc, Sharp Laboratories, ACAL Energy Ltd) to develop models/numerical codes for design purposes. He has received funding from the TSB, FP7, dstl and directly from industry.