# PINN-combined-with-LSTM

**Repository Path**: willeetony/PINN-combined-with-LSTM

## Basic Information

- **Project Name**: PINN-combined-with-LSTM
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2025-06-15
- **Last Updated**: 2025-06-15

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# PINN-combined-with-LSTM
Physics-informed neural network (PINN) has attracted significant attention for various fluid related problems in recent years. Although the basic principle allows the PINN to require less data to train a reliable model, data-driven PINN often suffers from the quantity and quality of the training data, without which will result in diminishing fitting capability in the edge regions and extrapolation beyond the training data. To address this issue, a new coupled model called LSTM-PINN (LP) is proposed, which combines the advantages of Long Short-Term Memory (LSTM) in handling long-term dependencies for time-sequenced data with the existing PINN framework. By incorporating the time-sequenced predictions at different spatial points generated by the LSTM into the training set of the PINN, the edge errors that refer to errors at boundary of space or time are reduced and the time-sequenced prediction capability of flow field is enhanced. The comparative study is conducted on the velocities along two directions predicted by the coupled model with those obtained from the benchmark PINN model, while the numerical solutions in the case of two dimensional flow around cylinder described by the Navier-Stokes (NS) equations are selected as training dataset. The results demonstrate that the proposed LP model improves the accuracy of prediction compared to the conventional PINN model, showing great potential for flow field reconstruction and time series prediction.