Sensitivity Analysis of Transport and Radiation in NeuralPlasmaODE for ITER Burning Plasmas

Achieving sustained thermonuclear fusion in tokamak reactors [1, 2] requires a precise understanding and control of burning plasma behavior, particularly under the high-power deuterium-tritium (D-T) conditions anticipated in ITER [3-5]. In these plasmas, interactions among energetic fusion alpha particles, electrons, and ions give rise to complex nonlinear processes, including collisional heating, radiative losses, impurity effects, and multi-region energy transport. These dynamics are strongly influenced by both global and local plasma parameters, such as magnetic field strength, safety factor, impurity concentration, and transport coefficients. Quantifying the sensitivity of plasma behavior to these parameters is essential for robust scenario design, performance optimization, and predictive control of fusion reactors. Previous studies [4, 6-10] have proposed various models to simulate burning plasma dynamics. However, many of these approaches rely on empirical scaling laws with limited flexibility across operational regimes. To address these challenges, NeuralPlasmaODE [11, 12] was developed as a data-informed, multi-region, multi-timescale modeling framework based on neural ordinary differential equations (Neural ODEs) [13, 14]. This framework builds upon prior nodal modeling of tokamak plasmas [10, 15-18] and has demonstrated strong performance in capturing energy transport and species interactions in both DIII-D and ITER scenarios. In this work, we extend NeuralPlasmaODE [11, 12] to conduct a comprehensive sensitivity analysis of transport and radiation effects in ITER burning plasmas.