更多精彩视频登陆网站

www.koushare.com

图 | Bastiaan J. Braams

题   目：Machine learning of equivariant functions inspired by atomistic modelling and three-dimensional image processing

报告人：Bastiaan J. Braams

单   位：Centrum Wiskunde & Informatica, Dalian Institute of Chemical Physics

时   间：2019-09-06

地   点：厦门大学化学化工学院

扫码观看精彩报告视频

报告摘要

In my talk I noted several works that use (finally) linear regression: the Gaussian Approximation Potential (Gábor Csányi et al.), Solid Harmonic Wavelet Scattering (Stéphane Mallat et al.) and the Atomic Cluster Expansion (Ralf Drautz). I also described recent work employing deep neural networks and there I emphasized two key aspects of the some of the recent work: the networks are structured and parameterized.

Structured: Relative to a standard feed-forward neural network, in these networks each "neuron" holds a structured value and the layer-to-layer connections are compatible with that structure. The concept is familiar from machine vision where the structured value is a two-dimensional array of real values (a pixel image in the input layer) and the compatible operation is a convolution with compact kernel. In our applications the structured value can be a feature vector (in the SchNet approach from Berlin) or also an object that has welldefined transformation properties under rotations of 3D space (in the equivariant neural networks from the Amsterdam group, the Tensor Field networks from the Google people and the more recent DeepPot work from Princeton).

个人简介

Bastiaan braams is an affiliated researcher in the Multiscale Dynamics group in DICP (December 2016 - present). Previously employed at International Atomic Energy Agency (IAEA), Vienna, Austria (2009-2016), Emory University, Atlanta, GA (2003-2009), New York University, New York City, NY (1989-2003), Princeton Plasma Physics Laboratory, Princeton, NJ (1986-1989) and on PhD research at FOM Institute for Plasma Physics (now DIFFER), Nieuwegein, Netherlands. The PhD work was done at Max Planck Institute for Plasma Physics, Garching, Germany and at UKAEA Culham Laboratory. Basic university education in Utrecht (theoretical physics) and Eindhoven (mathematics and computing science). Interested in scientific computing and data analysis with applications in plasma physics, molecular modelling, and atomic physics. His has a broad view of research interest, including fusion energy research, quantum chemistry and molecular modelling. 

会议简介

2019年9月3日-6日，由固体表面物理化学国家重点实验室（厦门大学）、福建省理论与计算化学重点实验室和厦门大学化学化工学院主办的“机器学习-计算化学Workshop”在厦门大学化学化工学院举办。本次Workshop邀请了相关领域的研究者报告领域前沿进展，并设置Hands-on tutorials环节帮助学员们熟悉代码的使用。此次Workshop的举办增进了不同领域研究者的交流，促进了开源共享的观念传递，希望推动大数据技术在计算化学和材料模拟等领域的应用。

●【学术视频】机器学习-计算化学Workshop | 麻省理工学院谢天：Tutorials for CGCNN and GDyNets

●【学术视频】机器学习-计算化学Workshop | 上海大学欧阳润海研究员：Data-Driven Materials Discovery with the Method SISSO

●【学术视频】第五届全国光孤子学术研讨会 | 佛蒙特大学杨建科教授：Optical PT Symmetry Theory, Applications and Generalizations

●【学术视频】第四届中国能源材料化学研讨会 | 中国科大、中科院大连化物所包信和院士：Journal of Energy Chemistry--with fast rising impact and awareness

●【大学频道】中国科学院大学卡弗里理论科学研究所呈献 | 西安交通大学刘博教授：Chiral orbital magnetism and topological supersolids in ultracold atoms

●【学术视频】ICFSI-17 | 耶拿大学Friedhelm Bechstedt: Validity of Anderson rule for interfaces between 2D semiconductors

●【学术视频】中国化学会第十一届全国有机化学学术会议 | 香港中文大学谢作伟院士：A Toolbox for Controlled Functionalization of Carboranes

●【学术视频】北京量子磁性暑期学校 | 中国科大邓友金教授：Clock Monte Carlo Methods & Lifted Worm Algorithms

●【学术视频】中国化学会第四届中国能源材料化学研讨会 | 南开大学周震教授：钠离子电池和混合电容器的研究进展

● 北大吴飙教授《简明量子力学》PDF电子教材免费下载

END