Machine Learning Prediction of Carbon Quantum Dots-Induced Crop Growth Promotion Comparing Solution and Film Formulations

Abstract

Machine learning enables accurate prediction of carbon quantum dots (CQDs)-induced crop growth promotion using a limited experimental dataset. Two distinct CQD formulations were evaluated: solution-type DIW-CQDs applied directly to seeds and membrane-type FA-CQDs@polyvinyl alcohol (PVA) films placed over seedlings, using pea and wheat as model crops. Among nine classical regression models systematically compared, the decision tree algorithm exhibited superior predictive performance, achieving a test set R² of 0.9878, RMSE of 10.34, and MAE of 2.84. Feature importance analysis revealed that chlorophyll a content is the dominant physiological indicator, contributing 94.69% to the prediction of crop vigor index. A simplified model employing only chlorophyll a as the sole input achieved comparable accuracy (R² = 0.9894), reducing the number of required input features from seven to one and substantially decreasing experimental measurement costs. This streamlined approach enables rapid, high-throughput screening without compromising predictive reliability. The established prediction framework supports three major application scenarios: concentration optimization for maximum growth promotion, formulation screening between solution and film types, and cross-species adaptation for different crops. By integrating material design, spectral engineering, and data-driven machine learning, this work provides a practical decision-making tool for the precision application of CQD-based formulations in nanoagriculture.