Machine Learning in Complex Organic Mixtures: Applying Domain Knowledge Allows for Meaningful Performance with Small...

Area of Science:

• Chemistry
• Environmental Science
• Data Science

Background:

• Quantifying components in complex mixtures is a persistent challenge across scientific disciplines.
• Machine learning (ML) has driven advances in "omics" fields but is underutilized in other chemistry areas due to small datasets.
• Environmental samples often yield limited data, hindering conventional ML applications.

Purpose of the Study:

• To develop an effective ML approach for analyzing complex mixtures using small environmental datasets.
• To demonstrate the impact of domain knowledge on ML model performance in data-scarce scenarios.

Main Methods:

• Utilized a small dataset of 35 high-resolution mass spectra from Canadian petroleum fractions.
• Applied machine learning models integrated with specific domain knowledge.
• Compared performance against unconstrained ML approaches.

Main Results:

• ML models incorporating domain knowledge achieved notable performance.
• The approach proved effective even with a limited number of spectra.
• Demonstrated successful analysis of complex mixtures from environmental samples.

Conclusions:

• Specific domain knowledge significantly enhances ML model performance for complex mixture analysis.
• This methodology offers a viable solution for analyzing small environmental datasets.
• The findings pave the way for broader ML adoption in environmental chemistry.