Physics-informed machine learning for TCO-layer thickness prediction and process analysis from multi-spectral images

We present a machine learning model for a robust and fast evaluation of thickness maps of Transparent Conducting Oxide (TCO) layers based on multichannel spectral images only. The model is applicable for the quality inspection of heterojunction solar cells with textured surfaces and an amorphous silicon layer stack beneath the TCO layer. Within our physics-informed approach, synthetic data are created online for model training by simulating reflection maps for given TCO thickness variations. The developed method determines a full-scale TCO-thickness map in 1 s from inline measurable RGB image data. The spatially resolved analysis allows inline quality inspection of the thickness distributions. Additionally, the thickness profile at the edges is inspected with high spatial resolution in Silicon heterojunction solar cell precursors, where TCO edge exclusion at the rear side is required to avoid shunting. We demonstrate our approach by quantifying the completeness and masking area for narrow masks, which is a process optimization step for increasing cell efficiency. We derive sorting criteria for an early-stage process control regarding shunts and quantify the influence of the positioning accuracy of the mask on the short-circuit current. • Physics-informed machine learning model trained purely with synthetic data. • Fast and full-scale thickness determination of TCO layers from inline measurements. • Quality inspection for heterojunction solar cells with textured surfaces. • Early-stage process control regarding edge exclusion properties.