Machine Learning Surrogate Models for Mechanistic Kinetics: Embedding Atom Balance and Positivity.

Area of Science:

• Computational chemistry
• Chemical kinetics modeling
• Multiscale simulations

Background:

• Multiscale simulations of reactive flows are vital but computationally expensive due to detailed chemical kinetics.
• Current surrogate models for reactive chemistry offer speedups but struggle with physical consistency.
• Machine learning models for chemical kinetics must ensure atom balance and positive concentrations.

Purpose of the Study:

• To develop a method that enforces atom balance and guarantees positive concentrations for machine learning models in chemical kinetics.
• To improve the physical consistency and reliability of surrogate models for reactive flows.

Main Methods:

• Introduction of a positivity preserving projection technique.
• Implementation of a correction by linear interpolation backtracking.
• Validation using atmospheric chemistry, heterogeneous catalysis, and synthetic reaction systems.

Main Results:

• The proposed method simultaneously guarantees atom balance and positivity of predicted concentrations.
• Demonstrated successful application in diverse chemical systems, including atmospheric chemistry and catalysis.
• Achieved exclusively positive model predictions that conform to atom balance without sacrificing accuracy.

Conclusions:

• The developed approach enhances the physical realism of machine learning-based chemical kinetics models.
• This method addresses key limitations in surrogate modeling for reactive flows.
• Enables more accurate and reliable multiscale simulations in various scientific fields.