From Lab to Field: Developing a Benchtop Raman Spectroscopy for Pesticide Detection in Food-2

This work package focuses on training an artificial intelligence algorithm using data from WP4 and WP6 (Raman spectra of various pesticides at various concentrations obtained using different nanofabricated surfaces) to classify pesticides and estimate quantities. Deep neural networks will form the basis for prediction, complemented by scalable machine learning algorithms to address potential issues with data interpretation and result scalability.
Data preparation will involve coding Raman wavelengths and intensities, subtracting control spectrum data, and background cleaning. Dimensionality reduction techniques like PCA and autoencoders will be employed to enhance the accuracy of AI algorithms. Band intensity normalization will improve pesticide quantity estimates.
Algorithm design will prioritize two criteria: Accuracy and Explainability. While neural networks are expected to provide the best accuracy, they lack explainability. Therefore, white-box models like Explainable Boosting Machines (EBM) will be used alongside neural networks to interpret which peak values influence specific pesticide predictions. The main success criteria are pesticide classification, which has a 90% accuracy rate, and pesticide quantity estimates, which have a 0.9 R2 score. Various algorithms and combinations will be tested throughout the project to meet these criteria.
We already have a PURE student working on this project, and we would like to accommodate one more student to assist the current student with the project.
Please only apply if you are responsible and can allocate the necessary time and effort to the project. We only continue with the self-disciplined and self-motivated.
References
https://www.lignonanoplatform.net/arastirma-programlari/arastirma-programi-4