报告人: Prof. Lei Wang （王磊）
Institute of Physics CAS（中国科学院物理研究所 ）
地点：江湾校区物理楼C311
时间: 9月5日周四 10:00-11:00
题目: Neural Canonical Transformation
摘要:We construct and train symplectic neural networks to perform canonical transformations for classical Hamiltonian systems. The neural canonical transformation changes the physical variables towards a latent representation with an independent harmonic oscillator Hamiltonian. As a consequence, the system's phase space density flows towards a factorized Gaussian distribution in the latent space. Since the canonical transformation preserves the Hamiltonian dynamics, the approach captures nonlinear collective modes in the learned latent representation. We discuss practical implementations of the symplectic neural network and present training schemes based on variational calculation or density estimation in the phase space. After demonstrating the abilities of neural canonical transformation on toy problems including two-dimensional ring potential and harmonic chain, we apply the approach to identify slow collective modes for realistic problems such as the alanine dipeptide and the MNIST dataset.
