Direct Estimation of Electric Field Distribution in Circular ECT Sensors Using Graph Convolutional Networks

Area of Science:

• Electrical Engineering
• Computational Imaging
• Applied Physics

Background:

• Electrical Capacitance Tomography (ECT) imaging requires accurate electric field estimation for its forward model.
• Traditional Finite Element Method (FEM) solvers are accurate but computationally expensive, hindering real-time ECT.
• Developing faster methods for electric field prediction is crucial for advancing ECT applications.

Purpose of the Study:

• To investigate the use of Graph Convolutional Networks (GCNs) for direct, one-step prediction of electric field distributions in ECT.
• To evaluate the physical fidelity of GCN-predicted electric fields by comparing derived capacitance matrices with FEM-based results.
• To demonstrate the feasibility of using learned approximators to replace computationally intensive traditional solvers in ECT.

Main Methods:

• A numerical model of a circular ECT sensor was used.
• Graph Convolutional Networks (GCNs) were trained on Finite Element Method (FEM) simulated data.
• The GCNs directly predicted 2D electric field maps for all excitation patterns.
• Capacitance matrices were computed using both GCN-predicted and FEM-based electric fields.

Main Results:

• GCNs accurately predicted 2D electric field distributions for the ECT sensor.
• Strong agreement was observed between GCN-derived and FEM-based capacitance matrices.
• The GCN approach demonstrated high physical fidelity compared to traditional FEM solvers.
• The study confirmed the feasibility of using GCNs as fast, learned approximators for ECT forward modeling.

Conclusions:

• Graph Convolutional Networks offer a viable and efficient alternative to traditional FEM solvers for ECT electric field prediction.
• The GCN-based approach enables direct, one-step prediction of electric fields, significantly reducing computational load.
• This proof-of-concept study paves the way for real-time ECT imaging and control applications.