Uncertainty-Informed Training Set Construction for Robust Extrapolation in Materials Property Prediction

Uncertainty-aware prediction of thermal expansion in complex oxides via LLM-curated literature data

Machine learning is increasingly used in materials discovery, but its practical application is often limited by the time required to construct structured experimental datasets and by the lack of reliable uncertainty estimates. We introduce an integrated workflow that combines document-grounded large language model (LLM) literature mining, human verification, and probabilistic modeling to enable uncertainty-aware prediction of the coefficient of thermal expansion (CTE) in complex oxides. Experimentally reported CTE values, compositions, and measurement temperature ranges are extracted from full-text articles using a document-grounded large language model and subsequently normalized and verified. Composition-derived descriptors and the reported temperature bounds are used as inputs to a probabilistic regression model that predicts both the expected CTE and a composition-dependent uncertainty, enabling prediction intervals for new compositions. On held-out tests, the model achieves competitive mean accuracy relative to deterministic baselines while producing uncertainty estimates that increase systematically for sparsely represented or chemically distinct compositions, enabling risk-aware screening and prioritization. The workflow supports comparative screening of compositions with targeted CTE behavior and helps guide experimental selection for detailed thermophysical characterization. This study illustrates how LLM-assisted literature curation can be combined with uncertainty-aware machine learning to construct property prediction workflows for materials systems where experimental data are sparse and primarily available in the literature.

Machine Learning Materials Properties by Encoding Orbital-Projected Density of States

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.

Tackling real-world crystal structure prediction from powder X-ray diffraction data

Crystal structure prediction (CSP) from powder diffraction data is a central challenge in materials chemistry. Machine learning (ML) models show promise, but most are trained on idealized simulated data, limiting reliability on real experiments. Here, we assess real-world behaviour using the previously published deCIFer model as an example of PXRD-conditioned generative CSP. deCIFer is an autoregressive transformer that conditions each step of structure generation on encoded PXRD data, guiding token-wise predictions of space group, lattice parameters, and atomic positions. Using controlled robustness tests, we quantify performance under realistic artefacts (noise, background, peak asymmetry, and Scherrer broadening) and introduce metrics for accuracy and predictive uncertainty. deCIFer adapts smoothly to signal distortions and improves over unconditioned baselines when diffraction features remain informative, while expressing appropriate uncertainty as the PXRD pattern becomes underdetermined. Experimental PXRD tests recover the known structures of Si and CeO2 and expose the expected limitations for lower-symmetry Fe2O3 and nanocrystalline CeO2. Overall, ML-based CSP is fundamentally limited by the information content of PXRD, but can accelerate expert workflows by rapidly generating chemically plausible candidates and quantifying uncertainty, making such models valuable human-in-the-loop tools for real-world structure determination.

MatUQ: a benchmark for uncertainty-aware out-of-distribution materials property prediction with graph neural networks

Reliable uncertainty quantification (UQ) for graph neural networks (GNNs) under out-of-distribution (OOD) shifts remains insufficiently characterized in materials discovery. Existing benchmarks based on random splits can overestimate model reliability by underrepresenting structural extrapolation challenges. Here we introduce MatUQ, a benchmark built on structure-aware Smooth Overlap of Atomic Positions Leave-One-Cluster-Out (SOAP-LOCO) splitting, together with a training protocol that combines Deep Evidential Regression (DER) with dropout regularization, for evaluating GNN reliability under structural distribution shifts. Through systematic experiments spanning six materials datasets, twelve GNN architectures, and eight UQ strategies, we find that predictive accuracy and uncertainty quality are distinct capabilities that tend to decouple under OOD evaluation, and that uncertainty-metric leadership is largely non-transferable across datasets and target properties. We further find that as training data become scarce, the optimal strategy shifts from evidential-containing hybrids toward pure ensemble variance across all evaluated architectures, with the architecture holding the distributional optimum shifting correspondingly. Standalone evidential regression rarely attains per-model optima and benefits from hybrid pairing only under specific combinations of data density, inductive bias, and target metric. Monte Carlo dropout is less effective as a standalone uncertainty estimator under the tested configurations but can contribute within hybrid schemes. Overall, MatUQ provides a more rigorous benchmark for assessing uncertainty-aware GNNs in OOD materials discovery.

Predicting dielectric constants of crystalline materials using explainable machine learning and composition-aware feature engineering

Differential Learning for Robust Prediction of Thermal Stability with Application to Energetic Materials

Predicting thermal stability during handling and storage is essential for the design of safe and reliable energetic materials. However, experimental measurements vary significantly across laboratories due to differences in protocols and analysis methods, making it difficult to train reliable predictive models. We address this challenge through differential learning. Rather than predicting absolute decomposition temperatures, we instead train message passing neural networks to predict relative differences between pairs of molecules. This approach reduces sensitivity to systematic experimental errors and achieves>85% accuracy in ranking compounds by thermal stability, outperforming conventional regression methods on the same heterogeneous dataset. To understand what drives these predictions, we compare neural network models with interpretable alternatives built from descriptors derived from ab initio calculations and cheminformatics software. This analysis identifies bond dissociation enthalpy as a key determinant of thermal stability rankings, providing further insight into the complex chemistry of thermal decomposition. The differential learning framework generalizes across model architectures, from graph neural networks to classical descriptor-based approaches. Our results demonstrate that learning relative properties rather than absolute values offers a practical solution for modeling noisy experimental data, with direct applications in materials design where thermal stability predictions inform safety protocols.