Deductive machine learning models for product identification.

Area of Science:

• Chemistry
• Machine Learning
• Computational Science

Background:

• Contemporary machine learning (ML) in chemistry primarily uses inductive learning from fixed features.
• Many chemical workflows, such as lab automation and spectral interpretation, require deductive reasoning.
• Underdetermined prediction scenarios and contradictory information necessitate deductive strategies.

Purpose of the Study:

• To develop and demonstrate a general strategy for creating ML models capable of deductive reasoning.
• To combine individual inductive ML models into a larger deductive network for chemical applications.
• To address the limitations of purely inductive ML in complex chemical prediction tasks.

Main Methods:

• Designed and trained ML models by integrating multiple inductive models into a deductive network.
• Demonstrated the strategy on the task of deducing reaction products from spectral mixtures.
• Utilized a large dataset of over 1.1 million simulated spectra for training.

Main Results:

• The developed models successfully deduced reaction products from spectral mixtures.
• Models could distinguish intended from unintended reaction outcomes and identify starting materials.
• Models showed strong performance on untrained tasks, including structural inference and predicting minor products.

Conclusions:

• The strategy of combining inductive models into deductive networks enables ML models to perform deduction in chemistry.
• This approach overcomes deductive bottlenecks in chemical problems, proving they are not insuperable for ML.
• The findings open new possibilities for ML applications in complex chemical reasoning and prediction.