Machine learning for quantitative and qualitative defect analysis in semiconductors using hyperspectral cathodoluminescence

My PhD research project addresses the challenge of efficiently characterising dislocations in semiconductors by developing a novel, machine-learning-based method for fast analysis of cathodoluminescence (CL) data. Dislocations disrupt the electronic properties of semiconductors, thereby diminishing device yield and performance. To mitigate these issues, accurately identifying and quantifying the density and type of dislocations is essential for quality control and performance optimisation in semiconductor manufacturing.

Existing methods rely on techniques like atomic force microscopy (AFM), which, while precise, are time-intensive and impractical for in-line inspection due to their inability to provide rapid defect analysis.

My proposed research aims to overcome these limitations by leveraging machine learning (ML) to develop a new approach for defect characterisation, specifically focusing on multidimensional CL data sets. CL microscopy, a technique that reveals optical and electronic properties of materials through the emission of light when exposed to electron beams, is particularly suitable for mapping semiconductor defects.

However, current industrial applications of CL rely solely on signal intensity, which only allows for the calculation of dislocation density without information on more detailed dislocation properties. By harnessing CL's hyperspectral imaging capabilities, the project intends to capture a rich dataset that includes detailed spectral and polarisation information, offering a multidimensional view of dislocations that traditional methods cannot currently match.

This high-dimensional data will provide insights into both dislocation types (edge or screw) and additional properties, such as the Burgers vector.

The research will use a multi-microscopy approach, combining data from various high-resolution imaging techniques, such as AFM, to generate a comprehensive dataset. The samples for this data collection will mainly be sourced from within the Cambridge Centre for Gallium Nitride. These diverse datasets will serve as the training foundation for machine learning models capable of accurately discerning dislocation characteristics in CL images.

By training ML algorithms on this multi-microscopy data, the project seeks to create models that can learn to identify dislocation types and properties, ultimately offering a rapid and automated means of defect characterisation. The trained ML models will then analyze hyperspectral CL data to detect and classify dislocations in a way that is faster and potentially more adaptable to in-line semiconductor manufacturing processes than current methodologies.

The anticipated outcome of this project is an advancement in semiconductor inspection technology, enabling manufacturers to monitor defect density and distribution more efficiently, thereby enhancing yield management and device reliability.

In summary, this research aims to provide a high-speed, robust, and scalable defect characterisation technology using machine learning and hyperspectral cathodoluminescence imaging. By addressing the limitations of current methodologies and integrating multi-microscopy data with advanced ML techniques, this project seeks to establish a new standard for rapid and accurate dislocation analysis in semiconductor materials, ultimately contributing to more efficient manufacturing and higher-quality semiconductor devices.