RotNet: A Rotationally Invariant Graph Neural Network for Quantum Mechanical Calculations

Area of Science:

• Computational chemistry
• Materials science
• Machine learning in science

Background:

• Deep learning models are increasingly used for predicting molecular properties.
• Traditional methods struggle with rotational invariance, a key physical law, leading to complexity and reduced accuracy.
• Models lacking rotational invariance generalize poorly to rotated molecular data.

Purpose of the Study:

• To develop a rotationally invariant graph neural network (RotNet) for accurate and accelerated quantum mechanical calculations.
• To address the generalization deficiency in deep learning models caused by molecular rotations.
• To maintain physical plausibility in machine learning predictions for chemical and biological applications.

Main Methods:

• Proposed RotNet, a graph neural network incorporating effective transformations for rotational invariance.
• Learned crucial distance and angular information directly from atomic coordinates.
• Benchmarked RotNet on protein fragments, electronic materials, and QM9 datasets.

Main Results:

• RotNet demonstrated superior performance compared to existing baseline models.
• The framework generalized effectively to spatial data with varying molecular orientations.
• Achieved high accuracy, improved efficiency, and fast convergence in quantum mechanical calculations.

Conclusions:

• RotNet successfully ensures rotational invariance, overcoming limitations of traditional methods.
• The model's strong generalization ability and performance indicate significant potential for scientific applications.
• RotNet can accelerate studies in protein dynamics simulation and materials engineering.