Data fusion of spectroscopic data for enhancing machine learning model performance

Developing accurate industrial prediction models for complex industrial and geological applications remains a significant challenge, particularly when relying on limited and disparate spectroscopic data. Traditional data fusion methods often fall short in effectively integrating complementary information across different spectral sources, limiting predictive performance. Complex-level ensemble fusion (CLF) is presented as a two-layer chemometric algorithm that jointly selects variables from concatenated mid-infrared (MIR) and Raman spectra with a genetic algorithm, projects them with partial least squares and stacks the latent variables into an XGBoost regressor, thereby capturing feature- and model-level complementarities in a single workflow. When benchmarked against single-source models and classical low-, mid-, and high-level data-fusion schemes, the CLF technique consistently demonstrated significantly improved predictive accuracy. Evaluated on paired Mid-Infrared (MIR) and Raman datasets from industrial lubricant additives and RRUFF minerals, CLF robustly outperformed established methodologies by effectively leveraging complementary spectral information. Mid-level fusion yielded no improvement, underscoring the need for supervised integration. These results constitute the first evidence that a stacked, complex-level scheme can surpass all established fusion levels on real-world spectroscopic regressions comprising fewer than one hundred samples and provide a transferable recipe for building more accurate and resilient soft sensors in quality-control and geochemical applications.