Enhancing deep learning-based field reconstruction with differentiable learning framework

Sample codes for training of the general bi-level differentiable learning framework that seamlessly integrates reconstruction models with sensor placement optimization(DSPO). The present framework co-optimize sensor placement and field reconstructions based on neural networks.

Reference

Xu Liu, Wei Peng, Xiaoya Zhang, Xiaoyu Zhao, Weien Zhou, Wen Yao, and Xiaoqian Chen, "Enhancing deep learning-based field reconstruction with differentiable learning framework,", arXiv preprint arXiv:.

Information

Author: Xu Liu This repository contains

1. cylinder2D case:
   • The training dataset file Cy_Taira.pickle in "cylinder2D/data" can be download in https://modelscope.cn/datasets/shanliwa/ReconstructionDataset/files
   • DSPO_fno_cy.py: train FNO with DSPO.
   • DSPO_gnn_cy.py: train GCN with DSPO.
   • DSPO_mlp_cnn_cy.py: train GCN with DSPO.
   • DSPO_mlp_cy.py: train mlp with DSPO.
   • DSPO_podnn_cy.py: train podnn with DSPO.
   • DSPO_senseiver_cy.py: train senseiver with DSPO.
   • CylinderDataset.py: load the cylinder2D data for training.
   • sensor_ini.py: load the init placement.
2. models file: reconstruction models including CNN, MLP, PODNN, GCN, FNO, Senseiver and so on.
3. utils file: files including differential operator.

Requirements

• python 3.X (>3.8)
• torch torch1.13.1+cu116
• numpy
• pandas
• pickle