利用物理引导的深度学习克服数据稀缺
机器学习
2026-01-12 v4 计算工程、金融与科学
摘要
深度学习（DL）严重依赖数据，数据的质量显著影响其性能。然而，在许多实际应用中，如结构风险评估和医学诊断，获取高质量且标注良好的数据集可能极具挑战性甚至不可能。这对深度学习在这些领域的实际应用构成了重大障碍。物理引导的深度学习（PGDL）是一种新型深度学习，能够将物理定律整合到神经网络训练中。这可应用于任何受物理定律控制或支配的系统，如力学、金融和医学应用。已有证明表明，借助物理定律提供的额外信息，PGDL在数据稀缺的情况下仍能实现极高的准确性和泛化能力。本综述对PGDL进行了详细考察，并对其在物理、工程和医学应用等各个领域中解决数据稀缺问题的应用提供了结构化概述。此外，本综述还指出了PGDL在数据稀缺方面面临的当前局限性与机遇，并对PGDL的未来前景进行了全面讨论。
引用
@article{arxiv.2211.15664,
title = {Utilising physics-guided deep learning to overcome data scarcity},
author = {Jinshuai Bai and Laith Alzubaidi and Qingxia Wang and Ellen Kuhl and Mohammed Bennamoun and Yuantong Gu},
journal= {arXiv preprint arXiv:2211.15664},
year = {2026}
}
备注
This submission has been withdrawn because the manuscript is a review-type work that has become substantially out of date. Further assessment indicated that several sections were incomplete and did not adequately reflect the current state of the literature. The authors therefore plan to develop a new manuscript and may submit it elsewhere