Speaker
Description
Exposing a surface to an ion beam or hot plasma leads to erosion and the development of surface structures on the nanoscale. Such nanostructures have been observed on tungsten samples exposed to plasma in the PSI-2 linear plasma device and the LHD stellarator. Existing studies show that these nanostructures can have an influence on the erosion process of plasma facing components (PFCs). Better understanding those can lead to improved erosion models and potentially refine simulations modelling the plasma wall interaction by including the effects of surface morphology on sputter yields and emission distributions. This could lead to more accurate predictions for the lifetimes of PFCs.
Simulating the evolution of surface structures during ion beam impact quite often suffers from large computational effort. A more convenient description of the evolution of these structures is possible using a Kuramoto-Sivashinsky (KS) type model whose parameters we aim to infer for given experimental data. For real world data, only a single surface profile is available for this time dependent chaotic process, because the surface’s height profile h(x,y) cannot be obtained in-situ, especially for experiments in a magnetic confinement device. There exist some previous approaches to this problem, for example training a regression model on the Fourier transform of the surfaces, or using large pretrained convolutional neural networks, finetuned on the regression task. We propose a different approach using the vision-transformer architecture and including additional physically informed input features to the training process. We show that training such a model on our KS-dataset leads to good predictive performance on unseen test data across different parameter regimes. Furthermore, we investigate the embedding the model creates for the profiles, by visualizing the so-called class-token of the transformer. We will present details of the method, parameter studies and results on our synthetic dataset.
The results show the capability of our architecture to understand and extract information from fusion relevant surface structures. This serves as a starting point for creating models that do not only predict analytical surface models, but that can be used as surrogates for simulations that create important input quantities for PWI codes. Our framework can be applied to this task by changing the target variables while keeping the architecture we have developed, as we have shown that it can learn surface structure dependent processes.
This work is part of the project FusKI, funded by the BMFTR under grant no. 13F1012C.