Similarity-Based Machine Learning for Small Data Sets: Predicting Biolubricant Base Oil Viscosities.

Area of Science:

• Computational chemistry
• Machine learning applications in chemistry

Background:

• Experimental chemical data is often scarce and expensive to acquire.
• Traditional machine learning methods struggle with limited datasets.
• Accurate prediction of molecular properties is crucial for chemical research and development.

Purpose of the Study:

• To develop a machine learning approach for accurate molecular property prediction using small datasets.
• To enhance prediction accuracy and reduce feature requirements compared to existing methods.
• To address the challenge of limited experimental data in chemical applications.

Main Methods:

• A similarity-based machine learning approach was developed.
• Molecules were grouped based on structural similarity using molecular fingerprints.
• Separate machine learning models were trained for each molecular group.
• The method was validated on dynamic viscosity and aqueous solubility datasets.

Main Results:

• The proposed method demonstrated comparable or superior performance to traditional approaches on larger datasets, using fewer features.
• Significant performance improvement was observed in predicting kinematic viscosity of biolubricant base oils (KV40) compared to transfer learning and Random Forest.
• The approach proved robust for limited data scenarios.

Conclusions:

• The similarity-based machine learning approach offers a robust framework for accurate molecular property prediction with limited data.
• This method is particularly effective when clear structural patterns exist within the dataset.
• The approach is generalizable to various molecular datasets, overcoming data scarcity challenges.