Exploring Important Features in Continuous Spectral Datasets Using Supervised Learning.

Area of Science:

• Materials Science
• Spectroscopy
• Machine Learning

Background:

• Spectral data analysis is complex, requiring domain expertise and time.
• Supervised machine learning (ML) can identify key features in labeled spectroscopic datasets.
• Understanding quantitative structure-property relationships (QSPR) is crucial for material science.

Purpose of the Study:

• To develop and evaluate ML models for analyzing spectroscopic data from ToF-SIMS and SAXS.
• To compare the performance of various ML algorithms in modeling spectral data and identifying important features.
• To investigate the impact of CNN architecture on analyzing position-invariant and non-invariant spectral data.

Main Methods:

• Regression ML models were trained on ToF-SIMS and SAXS spectroscopic data.
• Algorithms used include LASSO, PLS, Random Forest (RF), 1D-CNN, and MLP.
• Feature attribution measures and nested k-fold cross-validation were employed for performance evaluation.

Main Results:

• RF and 1D-CNN models demonstrated superior performance compared to linear methods (LASSO, PLS).
• MLP models underperformed relative to other tested algorithms.
• Feature attribution analysis provided insights into the drivers of spectral properties.

Conclusions:

• RF and 1D-CNN are effective for QSPR modeling using complex spectral data.
• The study highlights the utility of ML in extracting meaningful information from spectroscopic datasets.
• CNN architecture choices influence performance with invariant and non-invariant spectral data.