Spectral Graph Neural Networks Combined with Curriculum Learning: A New Paradigm for Predicting Molecular HOMO-LUMO Gaps

This article introduces a research study that combines spectral graph neural networks with curriculum learning to predict molecular HOMO-LUMO gaps. This method not only improves prediction accuracy but also significantly accelerates training speed, providing a new technical path for the fields of computational chemistry and drug discovery.

In the fields of computational chemistry and drug discovery, accurately predicting molecular electronic properties is a core challenge. The HOMO-LUMO gap is a key indicator for measuring molecular chemical activity and optical properties.

Traditional quantum chemistry calculations (such as DFT) have high accuracy but are costly, making large-scale screening unaffordable. Machine learning-based fast prediction models have thus become a research hotspot.

Molecules can be represented as graph structures. Traditional GNNs face limitations such as high homogeneity and difficulty in capturing global electronic properties.

Spectral graph neural networks are based on spectral decomposition of the graph Laplacian matrix and process graph signals in the frequency domain: low frequencies correspond to global topology, while high frequencies reflect local functional groups. Spectral convolution has a global receptive field, is insensitive to molecular size, and its features are interpretable.

Curriculum learning trains samples in an order based on difficulty: starting with simple ones and moving to complex ones.

This study defines difficulty by integrating molecular size, topological complexity, chemical diversity, and baseline error. It uses an adaptive scheduling approach: after meeting the standard, more difficult samples are introduced; if performance declines, it reverts to consolidate previous learning.

Synergistic effects of spectral graph networks + curriculum learning:

1. Training efficiency: The number of iterations is reduced by 30-40% at the same validation accuracy.
2. Prediction accuracy: Leading on standard datasets, with significant improvements in subsets of complex large molecules.

Reasons: Spectral graphs capture long-range interactions, and curriculum learning establishes basic patterns.

1. Drug discovery: Accelerate virtual screening and increase the number of molecules evaluated.
2. Material design: Guide the exploration of target gaps for organic electronic/photovoltaic materials.
3. Education: Open-source implementation provides AI-for-Science teaching resources for computational chemistry.

This study combines spectral graph neural networks with curriculum learning to provide an efficient and accurate path.

Limitations: High computational cost for ultra-large molecules. Future directions: Explore approximate spectral methods; expand to multi-molecule systems (such as solvation effects).