Structure-based out-of-distribution (OOD) materials property prediction: a benchmark study

npj Computational Materials, Journal Year: 2024, Volume and Issue: 10(1)

Published: July 4, 2024

Abstract In real-world materials research, machine learning (ML) models are usually expected to predict and discover novel exceptional that deviate from the known materials. It is thus a pressing question provide an objective evaluation of ML model performances in property prediction out-of-distribution (OOD) different training set. Traditional performance through random splitting dataset frequently results artificially high-performance assessments due inherent redundancy typical material datasets. Here we present comprehensive benchmark study structure-based graph neural networks (GNNs) for extrapolative OOD prediction. We formulate five categories problems three datasets MatBench study. Our extensive experiments show current state-of-the-art GNN algorithms significantly underperform tasks on average compared their baselines study, demonstrating crucial generalization gap realistic tasks. further examine latent physical spaces these identify sources CGCNN, ALIGNN, DeeperGATGNN’s more robust than those best (coGN coNGN) as case perovskites dataset, insights improve performance.

Language: Английский

---

When Do Quantum Mechanical Descriptors Help Graph Neural Networks to Predict Chemical Properties?

Journal of the American Chemical Society, Journal Year: 2024, Volume and Issue: 146(33), P. 23103 - 23120

Published: Aug. 6, 2024

Deep graph neural networks are extensively utilized to predict chemical reactivity and molecular properties. However, because of the complexity space, such models often have difficulty extrapolating beyond chemistry contained in training set. Augmenting model with quantum mechanical (QM) descriptors is anticipated improve its generalizability. obtaining QM requires CPU-intensive computational calculations. To identify when help properties, we conduct a systematic investigation impact atom, bond, on performance directed message passing (D-MPNNs) for predicting 16 The analysis surveys experimental targets, as well classification regression tasks, varied data set sizes from several hundred hundreds thousands points. Our results indicate that mostly beneficial D-MPNN small sets, provided correlate targets can be readily computed high accuracy. Otherwise, using add cost without benefit or even introduce unwanted noise degrade performance. Strategic integration unlocks potential physics-informed, data-efficient modeling some interpretability streamline

Language: Английский

---

Machine Learning Interatomic Potentials for Catalysis

Chemistry - A European Journal, Journal Year: 2024, Volume and Issue: 30(60)

Published: Aug. 7, 2024

Atomistic modeling can provide valuable insights into the design of novel heterogeneous catalysts as needed nowadays in areas of, e. g., chemistry, materials science, and biology. Classical force fields ab initio calculations have been widely adopted molecular simulations. However, these methods usually suffer from drawbacks either low accuracy or high cost. Recently, development machine learning interatomic potentials (MLIPs) has become more popular they tackle problems question deliver rather accurate results at significantly lower computational In this review, atomistic catalytic systems with aid MLIPs is discussed, showcasing recently developed MLIP models selected applications for systems. We also highlight best practices challenges give an outlook future works on field catalysis.

Language: Английский

---

Machine Learning in Solid‐State Hydrogen Storage Materials: Challenges and Perspectives

Advanced Materials, Journal Year: 2024, Volume and Issue: unknown

Published: Dec. 20, 2024

Abstract Machine learning (ML) has emerged as a pioneering tool in advancing the research application of high‐performance solid‐state hydrogen storage materials (HSMs). This review summarizes state‐of‐the‐art ML resolving crucial issues such low capacity and unfavorable de‐/hydrogenation cycling conditions. First, datasets, feature descriptors, prevalent models tailored for HSMs are described. Specific examples include successful titanium‐based, rare‐earth‐based, solid solution, magnesium‐based, complex HSMs, showcasing its role exploiting composition–structure–property relationships designing novel specific applications. One representative works is single‐phase Ti‐based HSM with superior cost‐effective comprehensive properties, to fuel cell feeding system at ambient temperature pressure through high‐throughput composition‐performance scanning. More importantly, this also identifies critically analyzes key challenges faced by domain, including poor data quality availability, balance between model interpretability accuracy, together feasible countermeasures suggested ameliorate these problems. In summary, work outlines roadmap enhancing ML's utilization research, promoting more efficient sustainable energy solutions.

Language: Английский

---

Advancing extrapolative predictions of material properties through learning to learn using extrapolative episodic training

Communications Materials, Journal Year: 2025, Volume and Issue: 6(1)

Published: Feb. 22, 2025

Language: Английский

---

A study of creep rupture life prediction for P91 steel with machine learning method: model selection and sensitivity analysis

International Journal of Pressure Vessels and Piping, Journal Year: 2025, Volume and Issue: unknown, P. 105494 - 105494

Published: March 1, 2025

Language: Английский

---

A generalizable framework of solution-guided machine learning with application to nanoindentation of free-standing thin films

Thin-Walled Structures, Journal Year: 2024, Volume and Issue: 200, P. 111984 - 111984

Published: May 7, 2024

Language: Английский

---

Impact of crystal structure symmetry in training datasets on GNN-based energy assessments for chemically disordered CsPbI3

Scientific Reports, Journal Year: 2025, Volume and Issue: 15(1)

Published: March 14, 2025

Robust solutions combining computational chemistry and data-driven approaches are in high demand various areas of materials science. For instance, such methods can use machine learning models trained on a limited dataset to make structure-to-property predictions over large search spaces. This paper examines the impact data selection mechanisms thermodynamic property assessments for chemically modified lead halide perovskite γ-CsPbI3 non-perovskite δ-CsPbI3. disordered states these phases, complete composition/configuration spaces built by adding Cd or Zn substitutions Pb Br I comprise 2,946,709 2,995,462 inequivalent spatial arrangements substituents, respectively. Using properties 1162 entries evaluated means density functional theory, we implement independent procedures training graph neural networks (GNNs). In each them, is constructed depending defect contents presence low- high-symmetry structures. The results show that symmetries structures significantly influence quality subsequent GNNs' result twofold increase errors due preferential

Language: Английский

---

When Do Quantum Mechanical Descriptors Help Graph Neural Networks Predict Chemical Properties?

Published: April 4, 2024

Deep graph neural networks are extensively utilized to predict chemical reactivity and molecular properties. However, because of the complexity space, such models often have difficulty extrapolating beyond chemistry contained in training set. Augmented model with quantum mechanical (QM) descriptors is anticipated improve its generalizability. obtaining QM requires CPU-intensive computational calculations. To identify when help properties, we conduct a systematic investigation impact atom, bond, on performance directed message passing (D-MPNNs) for predicting 16 The analysis surveys experimental targets, classification regression tasks, varied dataset sizes from several hundred hundreds thousands datapoints. Our results indicate that mostly beneficial D-MPNN small datasets, provided correlate well targets can be readily computed high accuracy. Otherwise, using add cost without benefit or even introduce unwanted noise degrade performance. Strategic integration unlocks potential physics-informed, data-efficient modeling some interpretability streamline de novo drug material designs. facilitate use machine learning workflows chemistry, provide set guidelines regarding how best leverage descriptors, high-throughput workflow compute them, an enhancement Chemprop, widely adopted open-source implementation property prediction.

Language: Английский

---

Modulating thermal and electrical conductivities in polymers: An approach toward extracting molecular design rules through atomistic simulations

Applied Physics Letters, Journal Year: 2024, Volume and Issue: 124(18)

Published: April 29, 2024

Polymers are extensively employed in diverse industries, including electrical equipment and electronic devices. Recent technological advancements have intensified the demand for dielectric polymers with both high insulation resistance thermal conductivity. We molecular dynamics simulations to clarify intricate relationship between structures, conductivity, ionic mobility from an atomistic point of view. Examined include polyethylene, polyvinyl alcohol, chloride, polyvinylidene fluoride, polytetrafluoroethylene, polychlorotrifluoroethylene, polyoxymethylene, polyethylene oxide. Based on elucidated correlations among force field parameters, we found that parameters can be clustered into four groups: group 1 (atomic bond constant angle), 2 (equilibrium angle dihedral 3 (side chain atom charges). Thermal conductivity showed relationships 1, correlation coefficients mostly exceeding 0.7 absolute value. Considering systematically altered within each computed mobility. When altering groups 2, a trade-off becomes evident. Conversely, increased while decreasing mobility, breaking relationship. The proposed clustered-parameter variation method predict changes through structure modifications. method, being general first-principles approach, is likely significant advantages design across range polymers.

Language: Английский

---