Learning a Local-Variable Model of Aromatic and Conjugated Systems.

Area of Science:

• Theoretical Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Deep learning (DL) offers new computational approaches for theoretical chemistry.
• DL aims to replace expensive ab initio quantum mechanics calculations with learned estimators.
• Questions arise about DL's ability to represent complex quantum chemical systems.

Purpose of the Study:

• To investigate the representability of nonlocal quantum chemical features using local-variable neural network models.
• To evaluate the efficacy of convolutional neural networks (CNNs) for modeling aromaticity and conjugation.
• To develop a novel DL architecture capable of efficiently capturing nonlocal interactions.

Main Methods:

• Analysis of convolutional neural network architectures for their limitations in representing nonlocal features.
• Development and implementation of a new wave-like neural network architecture.
• Testing the new architecture on aromatic and conjugated systems derived from molecule graphs.

Main Results:

• Convolutional architectures fail to efficiently represent nonlocal features like aromaticity and conjugation in large systems.
• The novel wave-like propagation architecture efficiently models these nonlocal features with high accuracy.
• The new architecture processes molecules in sublinear time with fewer parameters than CNNs.

Conclusions:

• Convolutional neural networks are insufficient for modeling certain nonlocal quantum chemical phenomena.
• Wave-like propagation architectures offer a computationally and representationally efficient solution for modeling complex quantum chemistry.
• Local-variable models can effectively represent some nonlocal features of quantum chemistry with appropriate architectures.