Machine-Learning Models for Accurate Prediction of Material Erosion of 3D Plasma-Facing Components in presence of RF sheaths

High-fidelity prediction of material erosion relies on computationally expensive plasma kinetic models (PIC, BCA, etc), primarily simulating Ion Energy-Angle Distributions (IEADs) of the particle impacting on the Plasma Facing Components. However, their application is constrained, especially in complex 3D tokamak geometries. This work introduces a new Machine Learning (ML) approach to solve the problem. We present a hybrid ML model merging convolutional neural networks and multi-layer perceptrons for efficient and precise IEAD emulation in presence of thermal and radio-frequency sheaths. Trained on a large database of simulations run by the hPIC2 code, our model can properly captures the nuances of diverse plasma conditions affecting IEADs. Predicting sputter yield, another critical erosion metric, is challenging due to traditional methods diverging from experimental data. We also introduce a Multi-Fidelity (MF) ML approach for sputter yield prediction, leveraging experimental data and simulations for improved regression. The integration of ML models was validated on a 3D mesh of the ICRH antenna of the WEST tokamak.