Analysis of Rutherford backscattering spectra with CNN-GRU mixture density network.

Area of Science:

• Materials Science
• Analytical Chemistry
• Physics

Background:

• Ion Beam Analysis (IBA) using MeV ion beams is crucial for determining surface elemental composition.
• High-throughput IBA mapping generates large datasets, overwhelming traditional analysis methods.
• Existing physics-based fitting algorithms are slow and can get stuck in local minima, hindering efficient data analysis.

Purpose of the Study:

• To develop a novel computational approach for analyzing complex IBA spectral data.
• To model the posterior distribution of Elemental Depth Profiles (EDP) using a Mixture Density Network (MDN).
• To improve the speed and accuracy of data analysis in IBA.

Main Methods:

• Implemented a Mixture Density Network (MDN) architecture.
• Utilized a Convolutional Neural Network-Gated Recurrent Unit (CNN-GRU) as the encoder module (EM).
• Employed a Multi-Layer Perceptron (MLP) for the Mixture Density Head (MDH) to model EDP from spectral data.

Main Results:

• The MDN approach performed comparably to the conventional Autofit method for simple and intermediate datasets.
• For complex datasets, the conventional Autofit method still showed superior performance.
• Combining the MDN with Autofit significantly improved accuracy and reduced computational time.

Conclusions:

• The developed MDN offers a viable alternative for analyzing IBA spectral data, particularly for less complex datasets.
• Integrating MDN with existing automatic fitting methods presents a powerful strategy for enhancing both accuracy and efficiency in IBA data analysis.
• This hybrid approach represents a significant advancement for accelerating and improving elemental composition analysis in IBA applications.