Outlier-Robust Surrogate Modeling of Ion-Solid Interaction Simulations.

Area of Science:

• Computational Physics
• Plasma Science
• Machine Learning Applications

Background:

• Complex plasma-wall interactions necessitate computationally intensive simulations.
• Large parameter spaces in simulations lead to poor coverage and unpredictable outliers.
• Robust analysis tools are crucial for exploring high-dimensional simulation data.

Purpose of the Study:

• To adapt Gaussian process (GP) methods for Bayesian adaptive exploration of plasma-wall interaction parameters.
• To enhance robustness against outliers using the Student-t process (TP) method.
• To investigate mixture likelihoods within a GP framework for outlier detection.

Main Methods:

• Restatement of Gaussian Process (GP) as a Bayesian adaptive exploration technique.
• Expansion of GP with Student-t Process (TP) for improved outlier robustness.
• Investigation of a Gaussian mixture likelihood within a GP model.

Main Results:

• Student-t process (TP) exhibits broader marginal probability distributions for hyperparameters in the presence of outliers compared to GP.
• The mixture likelihood model effectively describes both outlier and non-outlier behaviors.
• The proposed methods offer more robust surrogate surface establishment in complex parameter spaces.

Conclusions:

• Student-t process (TP) provides a more robust alternative to Gaussian process (GP) for analyzing data with outliers.
• Mixture likelihoods within GP offer a promising approach for handling heterogeneous data distributions.
• These advanced Bayesian methods improve the exploration and understanding of complex simulation data.