Analyzing Atomic Interactions in Molecules as Learned by Neural Networks.

Area of Science:

• Quantum Chemistry
• Computational Chemistry
• Materials Science

Background:

• Machine learning (ML) models achieve high accuracy in quantum chemistry predictions.
• Test set accuracy alone is insufficient for robust chemical modeling, particularly for molecular dynamics (MD).

Purpose of the Study:

• Develop a general analysis framework using explainable artificial intelligence (XAI) for atomic interactions in ML models.
• Evaluate how well ML models like SchNet and PaiNN learn physicochemical concepts.
• Identify ML model behaviors that lead to unstable MD simulations.

Main Methods:

• Applied XAI techniques to analyze atomic interactions within SchNet and PaiNN models.
• Compared ML-derived interactions against fundamental chemical principles.
• Investigated interaction strength, property prediction (intensive and extensive), and distance-dependent decay (many-body nature).

Main Results:

• ML models deviating from physical principles result in unstable MD trajectories, irrespective of high energy and force prediction accuracy.
• Analysis revealed insights into interaction strengths for different atomic species.
• The study highlighted issues with how models handle the polynomial decay of atomic interactions.

Conclusions:

• Explainable AI is crucial for assessing the physical realism of ML models beyond simple accuracy metrics.
• Current ML architectures may require modifications to better capture the polynomial decay of atomic interactions.
• Ensuring ML models adhere to chemical principles is vital for reliable molecular dynamics simulations.