摘要
由一名新闻记者兼机器人与机器学习的工作人员新闻编辑每日新闻-一项关于机器学习的新研究现在已经可用。根据NewsRx记者来自中国人民日报浙江的新闻报道，研究表明："地下水文过程的正向预测，以及需要重复评估正向模型的任务（如不确定性量化、反向建模、优化问题等）。深度学习基础D方法被越来越多地应用于促进正演建模过程（例如学习偏微分方程）和建立替代模型，以有效地替代基于物理的正演模型。我们的新闻编辑引用了东方理工学院的一篇研究，“然而，对于实际问题，可用的训练数据往往数量和质量有限，这严重影响了模型的学习能力。理论引导的机器学习(TGML)是近几十年来出现的，为了应对这种情况，通过引入数据预处理、神经网络设计或模型训练过程的理论指导，提高了模型的准确性、鲁棒性和鲁棒性。”本文综述了TGML在地下水文学问题中的最新研究进展,重点介绍了在模型训练过程中引入理论指导的三种方法,即在LOSS函数中加入附加物理规则的原理,以及基于TGML的前向预测代理模型、不确定性反演模型、不确定性反演模型和基于TGML的地下水文学模型,以及基于TGML的地下水文学模型。本文详细介绍了单-/两相流和污染物输运领域的反演建模和优化问题。
Abstract
By a News Reporter-Staff News Editor at Robotics & Machine Learning Daily News Daily News – A new study on Machine Learning is now available. According to news reporting originating from Zhejiang, People’s Repu blic of China, by NewsRx correspondents, research stated, “Forward prediction of subsurface hydrological processes, as well as tasks that require repetitive eva luation of forward models (e.g., uncertainty quantification, inverse modeling, o ptimization problems, etc.) can be computationally-intensive. Deep learning-base d approaches are increasingly applied to facilitate the forward modeling process (e.g., learning partial differential equations) and to build surrogate models a s efficient replacements of physics-based forward models.” Our news editors obtained a quote from the research from the Eastern Institute o f Technology, “However, for practical problems, the available training data are usually of limited quantity and quality, which significantly affects the models’ learning capability. Theory-guided machine-learning (TGML) emerged in the past decades to cope with such conditions. Through introducing theoretical guidance t o data preprocessing, neural network design, or model-training processes, the ac curacy, robustness, and generalizability of the trained model can be dramaticall y improved. Herein, a review is provided on the latest advances of TGML applied to subsurface hydrology problems. In particular, three ways of incorporating the oretical guidance into the model-training process are summarized, all of which a re based on the principle of adding additional physical regularizations to the l oss function. TGML-based surrogate models for forward prediction, uncertainty qu antification, inverse modeling, and optimization problems in the areas of single -/two-phase flow and contaminant transport are reviewed in detail.”
NETL
NSTL