Speaker
Description
Understanding the mechanisms of hydrogen recycling is a key factor in accurately predicting and managing plasma behavior in nuclear fusion reactors. In the present study, we develop a machine learning (ML) model capable of predicting the translational energy distributions and rovibrational states of hydrogen atoms and molecules emitted from tungsten plasma-facing components. These predictions play a critical role in evaluating the influence of recycled hydrogen on edge plasmas through neutral transport simulations.
The training dataset for the ML model[1] is generated using molecular dynamics (MD) simulations[2-4], which replicate hydrogen atom injection into hydrogen-saturated tungsten under various conditions, including different incident energies, material temperatures, and hydrogen-to-tungsten (H/W) ratios. These simulations provide detailed information on the resulting energy and rovibrational distributions.
To strike a balance between computational efficiency and predictive accuracy across a broad parameter space, we employ a fully connected neural network trained on 120 datasets derived from 24 distinct simulation scenarios, enhanced via random sampling for data augmentation. This ML model demonstrates high fidelity in reproducing the emission behavior of hydrogen species under monochromatic injection conditions.
The model is further generalized to more realistic plasma conditions by incorporating a shifted-Maxwellian distribution for the incident energy, accounting for the energy gain of ions in the sheath region. Two methodologies are proposed: (1) numerical integration of the monochromatic ML model over the shifted-Maxwellian distribution, and (2) development of a new ML model trained directly on pre-integrated data with four input parameters—H/W ratio, material temperature, and ion and electron temperatures. The analysis reveals that elevated electron temperatures enhance atomic hydrogen emission, while lower electron temperatures favor the release of molecular hydrogen.
References
[1] S. Saito, et al., “Machine Learning-Based Hydrogen Recycling Model for Predicting Rovibrational Distributions of Released Molecular Hydrogen on Tungsten Materials via Molecular Dynamics Simulations”, accepted to Nucl. Mater. Energy, (2025).
[2] S. Saito, et al., Contrib. Plasma Phys. e201900152 (2020).
[3] S. Saito, et al., Jpn. J. Appl. Phys. 60, SAAB08 (2021).
[4] S. Saito, et al., Nucl. Fusion 64 126067 (2024).