Application of Neural Networks for SPARC Exhaust Modelling

Managing divertor heat fluxes is a critical challenge for tokamaks including compact high-fields such as SPARC. To prevent plasma-facing component (PFC) damage SPARC must operate in a detached regime. While the SOLPS-ITER code is a powerful tool for simulating such conditions, predictive use is hindered by uncertainties in fueling physics and transport coefficients. Further, SPARC operation plans to control the upstream separatrix conditions to avoid large transients like type-I ELMs. Thus SPARC's operational strategy for empirical tuning of boundary parameters requires robust, preemptive modelling. In this work, we present a neural network-based approach to mapping the SPARC divertor exhaust operational space. To account for modeling uncertainties, we explore the full range of conditions that yield acceptable detachment, identifying empirical separatrix values for ELM/ELM-free operation and the margin to operational limits. We first apply the existing cross-machine SOLPS-NN model and compare it to SPARC SOLPS-ITER results. These results are then used to both fine-tune the pre-trained model and train new networks from scratch. Finally, we compare the resulting operational space predictions including a benchmark to the 1D X-Lengyel model, assessing their agreement and uncertainty. This approach enables rapid assessment of exhaust scenario robustness and informs plasma discharge decisions critical to safe operation in SPARC.