Analyzing Atomic Interactions in Molecules as Learned by Neural Networks

Malte Esders,Thomas Schnake,Jonas Lederer,Adil Kabylda,Grégoire Montavon,Alexandre Tkatchenko,Klaus-Robert Müller

2024-10-18

Abstract:While machine learning (ML) models have been able to achieve unprecedented accuracies across various prediction tasks in quantum chemistry, it is now apparent that accuracy on a test set alone is not a guarantee for robust chemical modeling such as stable molecular dynamics (MD). To go beyond accuracy, we use explainable artificial intelligence (XAI) techniques to develop a general analysis framework for atomic interactions and apply it to the SchNet and PaiNN neural network models. We compare these interactions with a set of fundamental chemical principles to understand how well the models have learned the underlying physicochemical concepts from the data. We focus on the strength of the interactions for different atomic species, how predictions for intensive and extensive quantum molecular properties are made, and analyze the decay and many-body nature of the interactions with interatomic distance. Models that deviate too far from known physical principles produce unstable MD trajectories, even when they have very high energy and force prediction accuracy. We also suggest further improvements to the ML architectures to better account for the polynomial decay of atomic interactions.

Chemical Physics

What problem does this paper attempt to address?

The problem this paper attempts to address is that while machine learning force fields (MLFFs) can achieve unprecedented accuracy in quantum chemistry prediction tasks, this accuracy does not guarantee stable molecular dynamics (MD) simulations. To go beyond simple test set accuracy, the authors use explainable artificial intelligence (XAI) techniques to develop a general framework for analyzing atomic interactions and apply it to the SchNet and PaiNN neural network models. By comparing these interactions with a set of fundamental chemical principles, the extent to which the models learn physical chemistry concepts from the data is investigated. Specifically, the authors focus on the following aspects:

1. **Interaction strength between different atom types**: Analyze how the interaction strength between different atom types affects the model's predictions.
2. **Prediction of intensive and extensive quantum molecular properties**: Explore how the models predict intensive and extensive properties and analyze their interaction ranges.
3. **Decay of interaction strength with distance**: Investigate whether the interaction strength follows a power-law decay, particularly comparing it with the decay pattern of London dispersion forces.
4. **Many-body effects**: Analyze the anisotropy of interaction strength, i.e., whether the interaction strength between atom pairs at the same distance varies due to the presence of neighboring atoms.

Through these analyses, the authors aim to reveal the performance of MLFFs in learning physical chemistry principles, thereby better understanding the models' prediction strategies and potential shortcomings. The ultimate goal is to improve the stability of MLFFs in molecular dynamics simulations, even in models with high energy and force prediction accuracy.