From the Thon Lab

Lee, H. J., Chiu, A., Lin, Y., Chintapalli, S., Kamal, S., Ji, E., & Thon, S. M. (2025). Predicting PbS colloidal quantum dot solar cell parameters using neural networks trained on experimental data. Advanced Intelligent Systems, 7(3), Article 2400310. https://doi.org/10.1002/aisy.202400310  

Context & Motivation 

Characterizing photovoltaic devices—and optimizing their performance—typically requires labor-intensive, multi-technique analysis. This is especially challenging for colloidal quantum dot (CQD) solar cells, where capturing detailed physical parameters often entails expensive and slow instrumentation workflows. Many machine learning (ML) models in this domain rely on simulated data, which limits their applicability because they may miss real-world nuances. 

What’s New: Neural Networks Trained on Experimental Data 

The authors introduce a set of neural networks trained on experimental measurements from PbS CQD thin-film solar cells: 

• Inputs: Illuminated current–voltage (J–V) curves collected across many spatial points on a single device. 

• Outputs: Predictions of complex materials parameters such as: 

• Carrier mobility 

• Relative photoluminescence (PL) intensity 

• Electronic trap-state density 

Experimental & Modeling Innovations 

• Spatially resolved sampling system: Instead of fabricating numerous devices, the team scans thousands of points across one solar cell using a custom setup—combining micrometer spatial resolution with device-scale coverage. 

• Neural network architecture: The models are based on ResNet-style convolutional blocks. A “neighborhood” approach integrates data from adjacent spatial points, capturing local spatial correlations and improving prediction accuracy. 

Key Results & Capabilities 

• Effective prediction of physical parameters from simple measurements (J–V curves), achieving practical estimation of mobility, PL intensity, and trap-state densities. 

• High-throughput capability: The approach transforms a single device into thousands of effective “micro-devices,” substantially increasing data density for ML training. 

• Spatial insight: The neighborhood-based ML captures spatial heterogeneity across the cell, revealing how local variation affects device behavior.ResearchGate 

Broader Implications 

• Accelerated characterization: The technique promises to streamline photovoltaic research workflows by reducing dependence on extensive instrumentation and multi-modal assays. 

• Generalizable strategy: Though demonstrated on PbS CQD solar cells, this method could be adapted to other optoelectronic devices, offering a scalable framework for rapid, data-driven materials insights. 

• Spatial diagnostics: The neighborhood modeling approach not only predicts key parameters but also supports spatial diagnostics—valuable for identifying performance heterogeneity or failure points within devices. 

Summary Table