Data-Driven Modeling and Initialization Strategies for Accelerating SOLPS-ITER

Edge plasma simulations using SOLPS-ITER are computationally expensive and sensitive to initial conditions, especially when exploring significant parametric variations or tailoring inputs to specific experimental scenarios. In this work, we develop a reduced modeling framework that combines two complementary strategies. First, we construct a KD-tree from a sparse database of converged SOLPS-ITER runs to identify the nearest-neighbor case in input space. The final state of that case is then used to initialize a new simulation, reducing CPU time by more than 30\% in most cases compared to cold starts. Second, we train two surrogate models: an ensemble of fully connected neural networks and a random forest regressor to emulate 1D radial profiles of plasma quantities, such as electron density and temperature, at the outer midplane and target. Both models are trained on a dataset of SOLPS outputs for DIII-D configurations, spanning variations in puff rate, core density, perpendicular diffusion, and diffusivity coefficients. The neural network ensemble achieves $R^2$ values greater than $0.8$ and low relative errors (<20\%) for most quantities. The combined approach enables faster scenario generation, reduces convergence failures, and lays the groundwork for real-time prediction and control applications in boundary plasma modeling.