A Machine Learning-Based Analysis for Efficient Predictive Modeling of Negative Hydrogen Ion Sources

Negative hydrogen ion sources (NHIS) are the preferred mode of plasma heating in tokamak devices due to superior neutralization efficiency. As such, accurate predictions of the H^- density are crucial for the design of such machines. This work exploits state-of-the-art statistical and data science tools developed for machine learning to reducing the computational cost of a global NHIS model while preserving its predictive capabilities. This model includes detailed pathways composed of hundreds of reactions to achieve high fidelity in reproducing the important kinetics features. Multiple approaches are deployed here with the aim of revealing which reactions are responsible for generating critical species that catalyze H^- production during the temporal evolution of the chemical system. Some of these techniques are grounded on graph theory and tackle the complex network of reactions based on dominant trajectories across the chemical graph. Unsupervised machine learning techniques are then applied for reducing the dimensionality by removing the uninformative reactions and clustering the species showing similarities in their dynamics. The result is a reduced order model, and a computationally inexpensive surrogate is constructed via probabilistic neural networks for emulating its steady-state solutions. The probabilistic attribute of this surrogate allows taking into account the uncertainties on rate coefficients while predicting the quantities of interest. Finally, non-linear manifold learning is used to study the global model's dynamical structure and introduce simplifications to the reaction set based on time scale analysis.