Scholarlink

<p>The integration of machine learning (ML) into modelling atmospheric and oceanic processes provides a robust alternative to traditional numerical methods, yet the performance of different State-of-the-Art ML models, particularly in the context of short- to medium-term forecasting, has not been thoroughly benchmarked. This study evaluates both standard and physics-informed neural network (PINN) models using gridded sea surface temperature (SST) data alongside six atmospheric predictors (cloud cover, relative humidity, solar radiation, surface pressure, u-component of velocity, and v-component of velocity) to explore spatial and temporal SST patterns. We assess four models—Convolutional Neural Network (CNN), Convolutional Neural Network combined with Long Short-Term Memory (CNN-LSTM), Transformer, and Transformer with Physics Informed Neural Network (PINN-Transformer)—across 7-day, 15-day, and 30-day lead times using Anomaly Correlation Coefficient (ACC), Nash-Sutcliffe Efficiency (NSE), Normalized Root Mean Square Error (NRMSE), and Mean Absolute Error (MAE). Our findings indicate that while CNN and CNN-LSTM models excel in short-term forecasts, the Transformer and PINN-Transformer models outperform at medium and long lead times, underscoring the significant advantage of integrating physical principles into ML frameworks. This integration not only enhances the models’ accuracy in longer-term predictions but also substantiates the crucial role of physics-informed approaches in advancing SST predictive capabilities. Such advancements are pivotal for improving climate models and enhancing the effectiveness of disaster management strategies, thereby reinforcing the essential role of hybrid modelling approaches in environmental science and operational forecasting.</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

Assessing machine learning and Physics-Informed models for Multi-Lead time sea surface temperature prediction in the Arabian sea

  • 期刊论文

  • 发表日期：

    2025年7月29日

查看全文
  • Akshay Sunil,
  • B. Deepthi,
  • Muhammed Rashid,
  • Rahul Sreedhar,
  • Gaurav Ganjir,
  • S. Adarsh

摘要

The integration of machine learning (ML) into modelling atmospheric and oceanic processes provides a robust alternative to traditional numerical methods, yet the performance of different State-of-the-Art ML models, particularly in the context of short- to medium-term forecasting, has not been thoroughly benchmarked. This study evaluates both standard and physics-informed neural network (PINN) models using gridded sea surface temperature (SST) data alongside six atmospheric predictors (cloud cover, relative humidity, solar radiation, surface pressure, u-component of velocity, and v-component of velocity) to explore spatial and temporal SST patterns. We assess four models—Convolutional Neural Network (CNN), Convolutional Neural Network combined with Long Short-Term Memory (CNN-LSTM), Transformer, and Transformer with Physics Informed Neural Network (PINN-Transformer)—across 7-day, 15-day, and 30-day lead times using Anomaly Correlation Coefficient (ACC), Nash-Sutcliffe Efficiency (NSE), Normalized Root Mean Square Error (NRMSE), and Mean Absolute Error (MAE). Our findings indicate that while CNN and CNN-LSTM models excel in short-term forecasts, the Transformer and PINN-Transformer models outperform at medium and long lead times, underscoring the significant advantage of integrating physical principles into ML frameworks. This integration not only enhances the models’ accuracy in longer-term predictions but also substantiates the crucial role of physics-informed approaches in advancing SST predictive capabilities. Such advancements are pivotal for improving climate models and enhancing the effectiveness of disaster management strategies, thereby reinforcing the essential role of hybrid modelling approaches in environmental science and operational forecasting.