{"title":"论基于神经网络的物理模拟器的不稳定性","authors":"Daniel Floryan","doi":"arxiv-2406.13101","DOIUrl":null,"url":null,"abstract":"When neural networks are trained from data to simulate the dynamics of\nphysical systems, they encounter a persistent challenge: the long-time dynamics\nthey produce are often unphysical or unstable. We analyze the origin of such\ninstabilities when learning linear dynamical systems, focusing on the training\ndynamics. We make several analytical findings which empirical observations\nsuggest extend to nonlinear dynamical systems. First, the rate of convergence\nof the training dynamics is uneven and depends on the distribution of energy in\nthe data. As a special case, the dynamics in directions where the data have no\nenergy cannot be learned. Second, in the unlearnable directions, the dynamics\nproduced by the neural network depend on the weight initialization, and common\nweight initialization schemes can produce unstable dynamics. Third, injecting\nsynthetic noise into the data during training adds damping to the training\ndynamics and can stabilize the learned simulator, though doing so undesirably\nbiases the learned dynamics. For each contributor to instability, we suggest\nmitigative strategies. We also highlight important differences between learning\ndiscrete-time and continuous-time dynamics, and discuss extensions to nonlinear\nsystems.","PeriodicalId":501167,"journal":{"name":"arXiv - PHYS - Chaotic Dynamics","volume":"85 1","pages":""},"PeriodicalIF":0.0000,"publicationDate":"2024-06-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"On instabilities in neural network-based physics simulators\",\"authors\":\"Daniel Floryan\",\"doi\":\"arxiv-2406.13101\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"When neural networks are trained from data to simulate the dynamics of\\nphysical systems, they encounter a persistent challenge: the long-time dynamics\\nthey produce are often unphysical or unstable. We analyze the origin of such\\ninstabilities when learning linear dynamical systems, focusing on the training\\ndynamics. We make several analytical findings which empirical observations\\nsuggest extend to nonlinear dynamical systems. First, the rate of convergence\\nof the training dynamics is uneven and depends on the distribution of energy in\\nthe data. As a special case, the dynamics in directions where the data have no\\nenergy cannot be learned. Second, in the unlearnable directions, the dynamics\\nproduced by the neural network depend on the weight initialization, and common\\nweight initialization schemes can produce unstable dynamics. Third, injecting\\nsynthetic noise into the data during training adds damping to the training\\ndynamics and can stabilize the learned simulator, though doing so undesirably\\nbiases the learned dynamics. For each contributor to instability, we suggest\\nmitigative strategies. We also highlight important differences between learning\\ndiscrete-time and continuous-time dynamics, and discuss extensions to nonlinear\\nsystems.\",\"PeriodicalId\":501167,\"journal\":{\"name\":\"arXiv - PHYS - Chaotic Dynamics\",\"volume\":\"85 1\",\"pages\":\"\"},\"PeriodicalIF\":0.0000,\"publicationDate\":\"2024-06-18\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"arXiv - PHYS - Chaotic Dynamics\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://doi.org/arxiv-2406.13101\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"\",\"JCRName\":\"\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"arXiv - PHYS - Chaotic Dynamics","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/arxiv-2406.13101","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
On instabilities in neural network-based physics simulators
When neural networks are trained from data to simulate the dynamics of
physical systems, they encounter a persistent challenge: the long-time dynamics
they produce are often unphysical or unstable. We analyze the origin of such
instabilities when learning linear dynamical systems, focusing on the training
dynamics. We make several analytical findings which empirical observations
suggest extend to nonlinear dynamical systems. First, the rate of convergence
of the training dynamics is uneven and depends on the distribution of energy in
the data. As a special case, the dynamics in directions where the data have no
energy cannot be learned. Second, in the unlearnable directions, the dynamics
produced by the neural network depend on the weight initialization, and common
weight initialization schemes can produce unstable dynamics. Third, injecting
synthetic noise into the data during training adds damping to the training
dynamics and can stabilize the learned simulator, though doing so undesirably
biases the learned dynamics. For each contributor to instability, we suggest
mitigative strategies. We also highlight important differences between learning
discrete-time and continuous-time dynamics, and discuss extensions to nonlinear
systems.