New ECT Benchmark Dataset Adds Electric Potential Maps for Physics-Guided ML
20,000 samples with full-field potential maps boost image reconstruction accuracy.
A new benchmark dataset from researchers Xinqi Zhang, Qiming Ma, and Lihui Peng at Tsinghua University explicitly incorporates electric potential field maps to improve deep learning-based image reconstruction in Electrical Capacitance Tomography (ECT). Traditional data-driven methods treat the ECT sensor as a black box, mapping capacitance to permittivity distribution directly, which ignores the underlying physics of the soft-field effect. This work introduces a COMSOL-MATLAB pipeline for an eight-electrode sensor that generates 20,000 randomized samples across four typical flow patterns, each including eight excitation-wise full-field electric potential maps alongside the usual capacitance vectors and permittivity images.
The paper includes evaluation protocols for both forward and inverse problems, demonstrating that the inclusion of electric potential maps significantly enhances reconstruction accuracy and robustness, particularly in out-of-distribution (OOD) scenarios. By making the latent physical field information explicit, the dataset lowers the barrier for integrating physics-guided machine learning into ECT. This standardized foundation could advance non-destructive testing and industrial process imaging, where ECT is commonly used.
- Dataset includes 20,000 COMSOL-MATLAB generated samples with full-field electric potential maps for each of eight excitation electrodes.
- Covers four typical flow patterns, enabling both in-distribution and out-of-distribution evaluation.
- Explicit potential field integration improves robustness and accuracy over black-box capacitance-to-permittivity models.
Why It Matters
Bridges the gap between deep learning and physical modeling, enabling more reliable ECT imaging for industrial diagnostics.