IISc researchers create machine learning model to predict semiconductor properties

Researchers at the Indian Institute of Science (IISc) and University College London have developed a new machine learning model to predict semiconductor properties using limited data. This model addresses the challenge of insufficient material property data, which hampers traditional methods. The new approach utilizes Graph Neural Networks and transfer learning, allowing the model to be pre-trained on a large dataset before being fine-tuned for specific tasks. This method has shown improved performance compared to models trained from scratch. The IISc team also introduced a framework called Multi-property Pre-Training, enabling the model to predict various material properties simultaneously. This advancement could enhance semiconductor manufacturing in India by predicting material behaviors more effectively.