Neural network-based surrogate model for 3D edge-plasma transport in the standard configuration of W7-X

This paper presents a neural-network surrogate model for Wendelstein 7-X (W7-X) edge
transport simulations, trained on an EMC3-EIRENE dataset that spans nearly the entire
operating-parameter space currently explored in the W7-X standard configuration. The model
uses an autoencoder to compress EMC3-EIRENE outputs into a low-dimensional latent space
representation, then a neural regressor maps EMC3-EIRENE input parameters to those latent
vectors, and finally both components are fine-tuned together to predict outputs directly from
inputs. In benchmark tests, the improved surrogate outperforms a traditional multilayer
perceptron model, most notably in predicting detached regime. Leave-one-value-out evaluation
indicates high accuracy within the interpolation domain, with minor degradation when
extrapolating. Relative to full EMC3-EIRENE runs, the surrogate provides over a 108 - fold
speedup, enabling large-scale parameter scans or real-time feedback control based on 3D
transport simulations to become feasible in the future.