Machine learning and deep learning enabled fuel sooting tendency prediction from molecular structure.

Area of Science:

• Chemical Engineering
• Computational Chemistry
• Materials Science

Background:

• Accurate soot formation modeling is vital for formulating advanced renewable fuels and achieving soot reduction.
• Machine learning (ML) and deep learning (DL) offer promising avenues for predicting the yield sooting index (YSI) from molecular structures.

Purpose of the Study:

• To evaluate and compare the performance of ML and DL models in predicting YSI from chemical structures.
• To propose a novel, tailor-made Convolutional Neural Network (CNN) architecture, SDSeries38, for regression tasks in fuel science.

Main Methods:

• Developed a novel quantitative structure-property relationship (QSPR) model for feature extraction and ML-based YSI prediction.
• Designed SDSeries38, a CNN with 9 feature learning modules and 1 regression module, for automated feature learning and regression.
• Compared the performance of the ML-QSPR model against SDSeries38 and classical CNNs.

Main Results:

• The ML-QSPR model demonstrated superior accuracy (RMSE = 7.563) compared to SDSeries38 (RMSE = 19.58).
• The ML-QSPR model exhibited faster computational speed and applicability to fuel mixtures.
• SDSeries38, while exceeding classical CNNs, highlights the need for specialized CNN architectures for regression tasks.

Conclusions:

• ML-QSPR models are highly effective for predicting YSI from molecular structures, offering advantages in accuracy, speed, and applicability to mixtures.
• Developing specialized CNN architectures for regression is crucial for advancing DL in this field.
• Modular CNN designs show promise for regression problems, with careful consideration of network depth to avoid vanishing gradients.