Ocean and Polar Research · 한국해양과학기술원

Article

A New Paradigm of Artificial Neural Network Learning for Atmospheric and Oceanic Predictions: Physics-Informed Neural Networks and Operator Learning
대기 및 해양 예측을 위한 인공신경망 학습의 새로운 패러다임: 물리정보 신경망과 연산자 학습

Dong-Hoon Kim, Il-Ju Moon
김동훈, 문일주

This study explores the potential of Physics-Informed Neural Networks (PINNs) and Operator Learning (OL) techniques in advancing scientific computation and prediction. PINNs … + READ MORE

This study explores the potential of Physics-Informed Neural Networks (PINNs) and Operator Learning (OL) techniques in advancing scientific computation and prediction. PINNs integrate physical principles into neural network training, enabling accurate modeling and prediction of complex physical systems, even in data-scarce and noisy environments. They have been successfully applied to solve challenging nonlinear partial differential equations (PDEs), such as Euler and Navier-Stokes equations, and continue to evolve with variants. Meanwhile, OL methods, represented by Deep Operator Network (DeepONet) and Fourier Neural Operator (FNO), focus on learning mappings between function spaces. These methods excel in high-dimensional data processing and have demonstrated remarkable performance in applications such as global atmospheric modeling with NVIDIA’s FourCastNet. Hybrid approaches, such as Physics-Informed DeepONet (PIDON) and Physics-Informed Neural Operator (PINO), combine the strengths of PINNs and OL. These methods leverage data-driven learning and physical constraints, achieving superior generalization and prediction accuracy. Notably, PINO enables zero-shot super-resolution predictions by integrating multi-resolution data with PDE constraints. While PINNs and OL individually present powerful tools for modeling and prediction, their computational cost and sensitivity to noise pose challenges. Hybrid approaches offer a pathway to address these issues by optimizing their integration through quantitative analyses. Future research directions include accelerating training through high-performance computing, extending applications to multiscale problems, and designing innovative loss functions to enhance data efficiency. This work synthesizes the latest advancements in PINNs, OL, and hybrid methods, providing a new paradigm for precise and efficient scientific computation across diverse fields. - COLLAPSE

PDF XML

References Information

21 February 2025