Using image-based machine learning and numerical simulation to predict pesticide inline mixing uniformity.

Area of Science:

• Agricultural Engineering
• Chemical Engineering
• Data Science

Background:

• Accurate pesticide inline mixing uniformity (PIMU) is crucial for direct nozzle injection systems (DNIS) to ensure system performance and optimize inline mixers.
• Supervised machine learning (ML) combined with computational fluid dynamics (CFD) simulations offers a novel approach for intelligent PIMU prediction using inline mixing images.

Purpose of the Study:

• To develop and evaluate an image-based ML approach for predicting PIMU in DNIS.
• To assess the effectiveness of different ML models and image processing techniques for accurate PIMU evaluation.

Main Methods:

• Principal Component Analysis (PCA) was used to reduce image data size while retaining essential information (98% information retention at <3% data size).
• CFD simulations provided reference values for ML models, validated by mixture sampling experiments.
• Enhanced images were used to train various ML models, including Generalized Linear Model (GLM), Support Vector Regression (SVR), BP-Neural Network (NNW), and Classification and Regression Trees (CART).

Main Results:

• PCA effectively reduced image dataset size for efficient analysis.
• NNW and CART models achieved prediction accuracies exceeding 97% for PIMU evaluation in both long-mixing tubes and jet mixers, outperforming GLM and SVR.
• Model performance was validated against CFD results and mixture sampling experiments, showing high accuracy and reliability.

Conclusions:

• PCA-enhanced, image-based ML provides a robust method for evaluating CFD-referenced PIMU in DNIS.
• NNW and CART models demonstrate superior performance and reliability for PIMU assessment compared to other tested models.
• These ML tools can significantly enhance the rationality and efficiency of PIMU evaluation in direct nozzle injection systems.