Synthetic spectra generated by boundary equilibrium generative adversarial networks and their applications with...

Area of Science:

• Spectroscopic analysis
• Computational chemistry
• Machine learning

Background:

• Limited calibration data is a major restriction in spectroscopic analysis, particularly for biological samples.
• Effective algorithms for simulating synthetic spectra from limited real spectra are lacking.

Purpose of the Study:

• To propose and evaluate a Boundary Equilibrium Generative Adversarial Network (BEGAN) for generating synthetic spectroscopic data.
• To assess the impact of diversity ratio on the quality and diversity of generated spectra.
• To improve predictive performance in spectroscopic analysis by using synthetic spectra to augment limited datasets.

Main Methods:

• Implementation of a Boundary Equilibrium Generative Adversarial Network (BEGAN) to generate synthetic spectra.
• Estimation of the diversity ratio's impact on spectral quality and diversity.
• Application of generated synthetic spectra in a Creating Diversity Partial Least Squares (CDPLS) consensus algorithm.

Main Results:

• BEGAN successfully generated synthetic spectra from two datasets.
• A negative correlation was observed between spectral quality and diversity.
• Synthetic spectra improved the predictive performance of the CDPLS algorithm.

Conclusions:

• BEGAN demonstrates potential for generating high-quality, derived homologous spectra.
• The method can effectively expand spectral datasets, particularly for small sample sets.
• This approach addresses limitations in spectroscopic calibration data for biological samples.