Abstract:Rapid advancements in machine learning (ML) are transforming materials science by significantly speeding up material property calculations. However, the proliferation of ML approaches has made it challenging for scientists to keep up with the most promising techniques. This paper presents an empirical study on Geometric Graph Neural Networks for 3D atomic systems, focusing on the impact of different (1) canonicalization methods, (2) graph creation strategies, and (3) auxiliary tasks, on performance, scalability and symmetry enforcement. Our findings and insights aim to guide researchers in selecting optimal modeling components for molecular modeling tasks.

What problem does this paper attempt to address?

The paper primarily explores the application of Geometric Graph Neural Networks (GNNs) in molecular modeling and evaluates the impact of different methods on model performance, scalability, and symmetry execution through empirical studies. Specifically, the paper attempts to address the following aspects: 1. **Comparison of Different Normalization Methods**: The researchers evaluated recently proposed normalization methods in ensuring or approximating Euclidean symmetry for handling 3D atomic systems. The experimental results indicate that approximate normalization methods (such as Stochastic Frame Averaging, SFA) can provide performance comparable to or even better than exact symmetry methods in certain practical application scenarios. 2. **Impact of Graph Construction Strategies**: The authors studied the impact of the graph construction process (i.e., how to create graphs based on the distances between atoms) and rewiring strategies on model performance. The results show that selecting an appropriate distance threshold is crucial when constructing geometric graphs, as too small or too large a threshold can affect model performance. Additionally, physics-inspired modules (such as Ewald-based methods) are effective for simple models like SchNet but do not significantly improve more complex models like FAENet. 3. **Role of Auxiliary Tasks**: The paper also explores how to leverage other tasks to improve FAENet's performance on specific tasks (such as IS2RE). Introducing Noisy Nodes as an auxiliary task can alleviate the over-smoothing problem and allow for the use of more interaction layers, thereby enhancing model performance. Furthermore, pre-trained models can improve performance across different tasks, demonstrating the knowledge transfer capability between atomic property prediction tasks. In summary, this paper aims to provide guidance to researchers, helping them choose the best modeling components based on specific task requirements to optimize the performance of geometric GNNs in molecular modeling tasks.

Improving Molecular Modeling with Geometric GNNs: an Empirical Study

Non-equilibrium molecular geometries in graph neural networks

3D Infomax improves GNNs for Molecular Property Prediction

Machine-learning accelerated geometry optimization in molecular simulation

A Review on the Applications of Graph Neural Networks in Materials Science at the Atomic Scale

TG-GNN: Transformer based Geometric enhancement graph neural network for molecular property prediction

Leveraging 2D molecular graph pretraining for improved 3D conformer generation with graph neural networks

Neural P$^3$M: A Long-Range Interaction Modeling Enhancer for Geometric GNNs

Molecular geometric deep learning

ChemRL-GEM: Geometry Enhanced Molecular Representation Learning for Property Prediction

Augmenting optimization-based molecular design with graph neural networks

A Hitchhiker's Guide to Geometric GNNs for 3D Atomic Systems

Geometric Deep Learning on Molecular Representations

3D Molecular Geometry Analysis with 2D Graphs

GemNet-OC: developing graph neural networks for large and diverse molecular simulation datasets

Molecule3D: A Benchmark for Predicting 3D Geometries from Molecular Graphs

Neural P^3M: A Long-Range Interaction Modeling Enhancer for Geometric GNNs

Molecular Graph Generation via Geometric Scattering

Geometry-Augmented Molecular Representation Learning for Property Prediction

Graph neural network framework for energy mapping of hybrid monte-carlo molecular dynamics simulations of Medium Entropy Alloys

Graph neural networks for materials science and chemistry