Interpretable machine learning prediction of biochar characteristics based on laser-Raman spectroscopy

Area of Science:

• * Materials Science and Engineering
• * Analytical Chemistry
• * Environmental Science and Engineering

Background:

• * Precise biochar characteristic detection is crucial for optimizing production and application.
• * Traditional methods are often time-consuming and labor-intensive.
• * Developing rapid and reliable prediction techniques is essential.

Purpose of the Study:

• * To develop interpretable machine learning (ML) models for predicting biochar characteristics using Raman spectroscopy.
• * To evaluate the performance of various ML algorithms (e.g., feedforward neural network, random forest) in predicting key properties.
• * To establish a framework for enhancing model interpretability and robustness.

Main Methods:

• * Application of Raman spectroscopy coupled with multiple ML models (extreme gradient boosting, support vector regression, feedforward neural network, random forest, ridge regression).
• * Development of a tripartite analytical framework integrating CARS, SHAP analysis, and mechanistic correlation for interpretability.
• * Model validation using enhanced datasets to assess robustness.

Main Results:

• * Feedforward neural network demonstrated superior predictive performance (R² = 0.89–0.95) for fixed carbon, volatile matter, H, O, and H/C, O/C ratios.
• * Highly accurate prediction of ash content (R² = 0.95) was achieved through integrated predictions.
• * The interpretable framework successfully linked spectral features to biochar structure.

Conclusions:

• * The combined approach of Raman spectroscopy and ML provides a rapid and reliable method for biochar characterization.
• * This methodology facilitates efficient control of biomass pyrolysis processes.
• * The study supports the development of online monitoring techniques for biochar production.