Evaluation of DNN and RNN for the determination of the chemical composition of dimensionality-reduced LIBS spectra of fusion-based materials

Laser Induced Breakdown Spectroscopy (LIBS) is often referred to as an in-situ and rapid analysis technique. Although the experimental setup is relatively simple, the quantification of elements in a sample containing multiple elements poses challenges for faster and reliable quantification. The application of machine learning (ML) techniques is one of the optimal solutions to achieve the quantified result in real-time. This study investigates the performance of Deep Neural Network (DNN), Gated Recurrent Unit (GRU), and Bidirectional Long Short-Term Memory (Bi-LSTM) models in analyzing the composition of first wall materials of thermonuclear reactors. The dataset was modelled based on 7400 simulated spectra at a resolution of 4000, each row comprising 41,730 data points. Initial evaluations revealed that GRU and Bi-LSTM models outperformed DNN in capturing spectral data relationships, as indicated by higher R² scores and lower Mean Squared Error (MSE). To mitigate computational complexity and eliminate redundant data, a bottleneck approach was used, which reduced the feature space to 1024 while enhancing predictive performance. Further enhancements were achieved through hyperparameter tuning using Polar Bear Optimizer (PBO), leading to significant improvements in the overall model accuracy. The integration of dimensionality reduction and hyperparameter optimization techniques demonstrated significant enhancement in the predictive capabilities of Recurrent Neural Network (RNN) models. This study emphasizes the potential of machine learning techniques in addressing the challenges associated with the rapid quantification of elements in complex fusion related samples.