Aug 2026· Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2· 0 citations· 19 references
Abstract
Machine learning (ML) has seen promising developments in materials science, yet its efficacy largely depends on detailed crystal structural data, which are often complex and hard to obtain, limiting their applicability in real-world material synthesis processes. An alternative, using compositional descriptors, offers a simpler approach by indicating the elemental ratios of compounds without detailed structural insights. However, accurately representing materials solely with compositional descriptors presents challenges due to polymorphism, where a single composition can correspond to various structural arrangements, creating ambiguities in its representation. To this end, we introduce PCRL, a novel approach that employs probabilistic modeling of composition to capture the diverse polymorphs from available structural information. Extensive evaluations on sixteen datasets demonstrate the effectiveness of PCRL in learning compositional representation, and analysis on model uncertainty highlights its potential applicability of PCRL in material discovery.
Machine learning models such as Crystal graph convolutional neural networks (CGCNN) have been widely adopted for the rapid and accurate prediction of crystalline material properties. However, these models suffer from a critical limitation: prediction errors increase sharply during extrapolation to out-of-distribution structures. In this study, we integrate Bayesian neural network (BNN) techniques into the CGCNN framework to quantify predictive uncertainty in formation energy estimation. Using a dataset of materials from the Materials Project, we assess the model’s extrapolation performance across varying lattice strains and diverse crystal systems. Our results reveal that CGCNN predictions are governed primarily by local atomic environments rather than by macroscopic crystal symmetry. Consequently, while the model extrapolates effectively to systems with similar local environments, it can exhibit overconfidence in low-symmetry structures where high errors occur despite low uncertainty. Building on these findings, we introduce a dataset construction strategy that utilizes BNN-derived uncertainty metrics while accounting for these architectural limitations in recognizing global structural changes. By strategically prioritizing the inclusion of structurally vulnerable regions identified through both uncertainty and structural analysis, the proposed methodology enables the development of models with improved extrapolation robustness and reliability in complex chemical spaces.
Machine learning force fields (MLFFs) combine the high accuracy of first-principles methods with the high efficiency of classical force fields, offering new opportunities for atomic-level studies of inorganic crystalline materials. We systematically summarize the research progress on MLFFs, elucidate their fundamental principles and developmental history, and categorically introduce the technical characteristics of representative models and relevant benchmarking platforms. We aim to review the advantages of MLFFs in overcoming traditional computational limitations across four domains: structural prediction and optimization, physical properties, defect and interface properties, and phase transitions and kinetic processes. The challenges of MLFFs are also examined in computational efficiency and simulation scale, accuracy and generalization ability, data requirements and training samples, model interpretability, and physical constraints, which offer a reference for the research and application of MLFFs in the field of inorganic crystalline materials.
Jing Yi, Yuxin Zhan, Yuanmao Hu et al.· Physical Chemistry, Chemical...· 0 citations
MOTIVATION
Protein conformation generation remains a fundamental challenge in structural biology and machine learning. Proteins are highly dynamic macromolecules, and their functions are governed not only by static 3D structures but also by their conformational flexibility. Therefore, generating a diverse ensemble of physically plausible structures is crucial for modeling conformational heterogeneity, characterizing intrinsically disordered proteins (IDPs), and supporting structure-based drug discovery. However, existing generative models often operate in continuous coordinate space. This approach makes it difficult to impose structural priors or handle complex, long-range dependencies, frequently leading to a significant trade-off where models must sacrifice structural validity to ensure conformational diversity.
RESULTS
To overcome these limitations, we propose a novel approach that learns hierarchical discrete representations of protein structures using vector quantization. Inspired by the success of VQ-VAE models in vision, we discretize residue-level structural contexts into learnable codebooks at multiple levels of granularity. Our framework first constructs a coarse-grained scaffold that captures global topology and secondary structures, and subsequently conditions on this scaffold to progressively refine the fine-grained local geometry. This coarse-to-fine generation mechanism enables both efficient sampling and highly accurate reconstruction. Through extensive experiments, we demonstrate that our method outperforms state-of-the-art models such as ESMDiff across challenging benchmark datasets, including BPTI MD trajectories and conformational-changing pairs (Apo/holo and fold-switch). Our model mitigates the trade-off between conformational diversity and structural validity, generating realistic ensembles while providing interpretable and scalable representations of protein geometry. Ultimately, this work offers a powerful new paradigm for flexible biomolecular modeling.
AVAILABILITY
The code is available at https://github.com/vanha9/hierarchical_conformation_generation. Archived at https://zenodo.org/records/21817083.
Seokjun On, Yujin Jeong, Kang-Hyeon Kim et al.· Bioinformatics· 0 citations
Graph neural networks have become the dominant machine-learning architecture for predicting materials properties from crystal structures. Yet the initialization of atomic node features has received comparatively little attention, and conventional approaches rely on static elemental descriptors that carry no information about the quantum-mechanical electronic environment of each atom in its crystalline host. Here we show that augmenting atomic node representations with site-projected orbital density of states (pDOS) fingerprints, computed directly from density functional theory calculations, yields systematic and substantial improvements in predictive performance.These representations are fused with Pettifor elemental embeddings at each atomic site before message passing. For the superconducting critical temperature $T_c$ and the optical dielectric constant $\epsilon_{\infty}$,the pDOS augmentation reduces prediction errors by 22.9% and 27.9%, respectively, relative to the elemental-descriptor baseline. These improvements are comparable to those achieved by doubling the training-set size. The gains are, however, contingent on training-set size. For the magnetic exchange energies of Heusler compounds, a substantially smaller dataset, the improvement is reduced,indicating that pDOS augmentation is most effective when the training data exceeds the length of the pDOS feature vector. We introduce an interpretable spectral attention-gating mechanism that reveals that the model autonomously learns to prioritize the orbital channels and energy windows most physically relevant to each target property. These results establish pDOS-augmented graph nodes as a broadly applicable strategy for infusing first-principles electronic-structure knowledge into graph networks, opening a practical route to high-accuracy property prediction in data-scarce regimes.
Paulo R. Pires, Pierre-Paul De Breuck, Mauro Fava et al.· 0 citations
Generative machine learning is increasingly used for inorganic crystal structure generation. Most models and the corresponding evaluation approaches rely on simple forms of crystal structure representation. In this paper, we showcase the power of atom-averaged features from pretrained Machine-Learning Interatomic Potentials (MLIPs), such as MACE, for such tasks. We first introduce a distance measure that assesses the output of material generative models by capturing both quality and novelty in a single distribution-based evaluation framework. In particular, we introduce the Coarse-Fine Transport Distance (CFTD) using two different featurizers, where the quality component is based on coarse MACE features. We showcase CFTD's versatility in capturing crystal-structure quality while also detecting memorization, and compare it with the recently introduced continuous SUN metrics. We further show that coarse MACE features can be used as guidance for a material generative model.
P. Hagemann, Katharina Ueltzen, Simon Müller et al.· 0 citations
The ElemeNet software package enables the training of advanced ML models for diverse properties and datasets with an enlarged range of elemental compositions, and introduces moiety predictions, a unified, general-purpose software package for molecular machine learning.
Jacob W. Toney, S. Darouich, Yiran Wang et al.· arXiv.org· 0 citations