Multi-task learning on microscopic hyperspectral data enables accurate classification of graphene oxide films

Area of Science:

• Materials Science
• Nanotechnology
• Data Science

Background:

• Precise characterization of graphene oxide (GO) is crucial for advanced applications but challenged by heterogeneity.
• Existing methods like AFM are too slow, and Raman spectroscopy is affected by GO's disorder.

Purpose of the Study:

• To develop a novel, high-throughput framework for GO characterization.
• To integrate microscopic hyperspectral imaging (mHSI) with multi-task learning (MTL) deep neural networks.

Main Methods:

• Utilized mHSI to capture detailed spectral information from GO films.
• Developed and trained an MTL deep learning model for GO classification into 3 primary and 12 thickness-based categories.
• Compared MTL model performance against a single-task learning (STL) model and pseudo-RGB image analysis.

Main Results:

• Achieved 97.1% classification accuracy for the 12-class GO thickness task on an independent test set.
• The MTL model demonstrated over three times lower error rate compared to the STL model.
• Hyperspectral data was critical, with accuracy dropping to 69.4% when using pseudo-RGB images.

Conclusions:

• Established a robust, data-driven paradigm for materials characterization.
• The mHSI-MTL framework offers a scalable, non-destructive solution for GO quality control.
• Enables advancements in fundamental research and high-throughput production of GO-based devices.