Electron flow matching for generative reaction mechanism prediction.

Area of Science:

• Computational Chemistry
• Machine Learning
• Chemical Reactivity

Background:

• Mass conservation is crucial for chemical consistency and reaction design.
• Current data-driven models for reaction prediction often fail to conserve mass.
• This leads to physically inconsistent predictions and 'hallucinatory' failure modes.

Purpose of the Study:

• To develop a deep generative model for chemical reaction prediction that explicitly enforces mass and electron conservation.
• To address limitations of existing models in terms of physical consistency and mechanistic interpretability.
• To improve the accuracy and generalizability of data-driven reaction outcome prediction.

Main Methods:

• Recasting reaction prediction as electron redistribution using flow matching.
• Employing a bond-electron (BE) matrix representation to ensure mass and electron conservation.
• Developing the FlowER model, a deep generative framework.

Main Results:

• FlowER enforces exact mass conservation, resolving common failure modes in reaction prediction.
• The model recovers mechanistic reaction sequences for novel substrates.
• It generalizes effectively to new reaction classes with minimal data through fine-tuning.
• FlowER enables estimation of thermodynamic or kinetic feasibility.

Conclusions:

• FlowER represents a significant advancement in data-driven reaction prediction by integrating fundamental chemical principles.
• The model bridges the gap between predictive accuracy and mechanistic understanding.
• This interpretable framework enhances chemical intuition in computational models.