Source-linked AI summary
Graph neural networks for materials science and chemistry
Patrick Reiser, Marlen Neubert, André Eberhard, Luca Torresi, Chen Zhou, Chen Shao, Houssam Metni, Clint van Hoesel, Henrik Schopmans, Timo Sommer, Pascal Friederich
TL;DR
Chemistry and materials science need machine-learning methods that can use molecular and materials structure for prediction, simulation, design, and synthesis-related tasks. This review explains GNN principles, representations, architectures, datasets, and applications, then outlines future development needs. It surveys evidence including data-efficient interatomic potentials and structure-based materials screening, while highlighting benchmark, distribution, and geometric-representation limitations.
Problem
Chemistry and materials tasks often require informative molecular or materials representations, while conventional models typically depend on predefined features rather than learning directly from full structure.
Method
The review synthesizes GNN formalism, structure representations, architectures, datasets, generative methods, and applications across chemistry and materials science.
Results
Across surveyed applications, GNNs support data-efficient interatomic potentials and structure-based screening of materials and catalysts.
Takeaways & Limitations
GNNs provide a structure-based machine-learning framework for materials-property prediction, simulations, inverse design, and broader materials-development workflows.
Takeaways & Limitations
Transfer to new materials domains remains constrained by limited benchmarks, distribution mismatch, and the need for geometric information for many quantum-mechanical properties.
Abstract
from arXiv · showhide
Machine learning plays an increasingly important role in many areas of chemistry and materials science, e.g. to predict materials properties, to accelerate simulations, to design new materials, and to predict synthesis routes of new materials. Graph neural networks (GNNs) are one of the fastest growing classes of machine learning models. They are of particular relevance for chemistry and materials science, as they directly work on a graph or structural representation of molecules and materials and therefore have full access to all relevant information required to characterize materials. In this review article, we provide an overview of the basic principles of GNNs, widely used datasets, and state-of-the-art architectures, followed by a discussion of a wide range of recent applications of GNNs in chemistry and materials science, and concluding with a road-map for the further development and application of GNNs.
1 Introduction
Machine learning is becoming integral to the natural sciences and is being applied across the materials-development cycle. GNNs extend this toolbox by learning from graph-structured representations relevant to chemistry and materials science.
- Machine learning now supports materials development from candidate discovery and property prediction through simulation, synthesis planning, and experimental analysis.
- GNNs operate on graph-structured data and are closely connected to geometric deep learning.
- GNNs can learn internal materials representations for task-specific property prediction, complementing or replacing hand-crafted features.
- The review introduces GNN formalism, compares architectures and benchmark datasets, surveys applications, and discusses inverse design, open challenges, and future developments.
2 Graph neural networks in materials science and chemistry
GNNs represent molecules and materials as graphs and learn task-specific embeddings through message passing. This section reviews their representations, architectures, benchmarks, applications, and open challenges in chemistry and materials science.
- Basic principles: Graphs represent atoms or entities as nodes and their connections as edges, supporting graph-, node-, and edge-level prediction tasks.
- Structure representation: GNNs learn molecular representations directly from full structures or geometries, reducing reliance on predefined feature representations.Learned node embeddings represent each atom together with its chemical environment.
- Basic principles: MPNNs propagate node information through neighboring edges, repeatedly updating embeddings to incorporate increasingly distant chemical environments.Repeated steps can reach K-hop neighborhoods, but long-range information may suffer from over-squashing or over-smoothing.
- Structure representation: Chemical graphs are useful for molecules and compounds, but quantum-mechanical and electronic-structure properties often require explicit geometric information.The review discusses geometric descriptors and architectures designed to incorporate atomic positions and distances.
- Architectures, benchmarks, and applications: The review compares architectures and benchmark datasets, while noting that model categories overlap and benchmark performance depends on splits, preprocessing, losses, and hyperparameters.It also covers graph generation, whose application to crystalline materials remains under-explored because reliable graph representations are difficult.
3 Applications
The review surveys GNN applications across chemistry and materials science, spanning property prediction, accelerated atomistic simulations, and reactivity and synthesis-route prediction.
- GNN applications range from predicting molecular and materials properties to accelerating atomistic simulations.
- The review also covers GNN-based prediction of molecular and materials reactivity and synthesis routes.
3.1 Molecular systems
GNNs support molecular property prediction, synthesis planning, and diverse chemistry applications by operating on molecular graphs, while also enabling faster and more scalable molecular dynamics simulations. Recent architectures improve physical representation, data efficiency, and transferability, but larger systems and longer timescales remain challenging.
- Molecular property prediction: GNNs address molecular property prediction without requiring predefined molecular feature vectors or fingerprints.Applications include regression and classification of molecular properties, including ADMET, electronic, optical, and toxicity-related properties.
- Molecular applications: GNNs have been applied across molecular-material domains, including drug discovery, bee toxicity, fuel ignition quality, and material-phase classification.Figure 2 summarizes applications including ADMET prediction, environmental effects, toxicity, and inverse molecular design.
- Molecular dynamics: GNNs and related machine-learning potentials replace costly ab initio energy-and-force calculations to accelerate molecular dynamics while aiming to retain accuracy.Molecular dynamics can address equilibrium, excited-state, non-equilibrium, slow, and rare-event processes, but current challenges include scale, timescale, data efficiency, generalization, transferability, and accuracy.
- Molecular dynamics: Equivariant models such as NequIP improve predictive performance and data efficiency by encoding geometric information through symmetry-aware message passing.On MD17, NequIP predicted energies and forces with higher accuracy than existing models while requiring up to three orders of magnitude less training data; it was also trained on CCSD(T) data.
- Scalable simulation and design: GNNs extend molecular simulations through large-scale and coarse-grained modeling, while specialized potentials support transferable excited-state simulations.Large biological systems involve far more atoms and longer configuration timescales than typical molecular-dynamics steps; DSGPM produced coarse-grained mappings nearly indistinguishable from human annotations, and DANN was transferable across azobenzene derivatives.
- Synthesis planning: Synthesis prediction and retrosynthesis remain difficult because chemical space is vast and currently demands trained chemists’ skills and experience.Machine-learning approaches including sequence models, transformers, and graph-based molecular embeddings aim to reduce manual effort in these tasks.
3.2 Crystalline and solid state systems
GNNs support crystalline and solid-state materials tasks spanning property prediction, disorder-aware modeling, simulation, synthesis prediction, and inverse design. Applications demonstrate broad utility, while active learning and improved datasets address generalization and transferability challenges.
- Materials property prediction and design: GNNs screen crystal databases for properties including adsorption volumes, stability, and functional behavior, reducing reliance on exhaustive conventional simulations.Crystal GNNs have been applied to methane adsorption in MOFs, convex-hull-distance prediction, and large-scale hypothetical-material screening.
- Materials property prediction and design: Domain-informed pooling improved methane-adsorption prediction in MOFs, while alloy-catalyst screening approached baselines trained on fully relaxed structures using only bulk-relaxed inputs.The latter removes the costly DFT relaxation of adsorbates from the screening process.
- Materials property prediction and design: Chemical-formula graphs yielded embeddings with better sample efficiency than other structure-agnostic approaches, extending GNN representation learning beyond explicit atomic structures.Each element is represented as a node weighted by its fraction in the formula.
- Challenges and future directions: Active learning can iteratively expand training data using uncertainty quantification, but broad transfer across dense, porous, and amorphous materials remains constrained by limited generic benchmark datasets.The review identifies dataset scarcity for crystalline, partially disordered, and amorphous structures as an important limitation.
- Disordered systems and defects: Disorder-aware GNNs represent doped or fractionally occupied sites using weighted elemental embeddings, and predictions from undoped training data usually matched DFT trends for doped structures.The approach also supports multi-fidelity training by encoding the DFT level in the model’s global state.
- Other applications and inverse design: GNNs have been applied to glass-phase classification, crystal dynamics, solid-state synthesis prediction, and periodic-graph generation for inverse crystal design.Glass classifiers outperformed CNNs, synthesis pipelines combine precursor and process representations, and generative approaches decode or relax new periodic structures.
4 Outlook
GNNs have advanced rapidly across materials simulation, screening, design, and synthesis, but their broader impact is limited by implementation, benchmarking, transfer, data, reliability, and interpretability challenges.
- GNNs have been successfully applied to materials simulation, screening, design, and synthesis, yet substantial open questions remain before their full potential can be realized.
- GNNs remain expert tools because they are difficult to implement, adapt, transfer, train, and apply across datasets and applications.Existing libraries provide state-of-the-art layers and models but often require expertise beyond conventional machine learning workflows.
- Reliable quantitative comparison requires common benchmarks, open training workflows, and open-source code, while hyperparameter optimization remains computationally expensive.
- Transfer from benchmark datasets such as QM9 is constrained because real applications involve larger, more diverse, and less densely sampled molecular or materials spaces.
- Materials applications are limited by fewer field-specific models and benchmarks, especially large labeled datasets for disordered and amorphous systems.This constrains transfer to porous, amorphous, and other materials settings despite GNNs’ suitability for representing disorder.
- GNN-based generative design still faces reliability and transferability challenges, with few convincing examples of machine-learning-designed materials.More interpretable models could help reveal correlations and causal relations in complex datasets.