Scholarlink

<p>Modeling physical systems with deep learning-based methods has received increasing attention, driven by promising results in domains such as weather forecasting, solid and fluid mechanics, chemistry or biology. This emerging paradigm, often referred to as <i>scientific machine learning</i>, encompasses a wide range of applications, each with distinct methodologies and best practices. In this work, we review the core components of physics-based deep learning with a focus on supervised learning for surrogate modeling, where deep learning models are trained to approximate the input-output behavior of physical systems using labeled data and, when available, additional physical knowledge. Rather than providing an exhaustive survey, we adopt a practitioner-oriented perspective and identify the key factors that govern model design and training. In particular, we propose a framework to guide the selection of appropriate architectures and training strategies, based on the characteristics of the target problem and the available computational resources.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

How to Learn Physics: A Practical Review of Scientific Machine Learning

  • 期刊论文

  • 发表日期：

    2026年7月14日

查看全文
  • Marien Chenaud,
  • José Alves,
  • Frédéric Magoulès

摘要

Modeling physical systems with deep learning-based methods has received increasing attention, driven by promising results in domains such as weather forecasting, solid and fluid mechanics, chemistry or biology. This emerging paradigm, often referred to as scientific machine learning, encompasses a wide range of applications, each with distinct methodologies and best practices. In this work, we review the core components of physics-based deep learning with a focus on supervised learning for surrogate modeling, where deep learning models are trained to approximate the input-output behavior of physical systems using labeled data and, when available, additional physical knowledge. Rather than providing an exhaustive survey, we adopt a practitioner-oriented perspective and identify the key factors that govern model design and training. In particular, we propose a framework to guide the selection of appropriate architectures and training strategies, based on the characteristics of the target problem and the available computational resources.