Derivative-based pre-training of graph neural networks for materials property predictions
General pre-training strategy of graph neural networks for materials science.
36 MATERIALS SCIENCE↗
Engineering topics
Fung, Victor
Publications and source records attributed to Fung, Victor.
Toward Quantum Chemical Free Energy Simulations of Platinum Nanoparticles on Titania Support
Platinum nanoparticles (Pt-NPs) supported on titania surfaces are costly but indispensable heterogeneous catalysts because of their highly effective and selective catalytic properties. Therefore, it is vital to understand their physicochemical processes during catalysis to optimize their use and to further develop better catalysts. However, simulating these dynamic processes is challenging due to the need for a reliable quantum chemical method to describe chemical bond breaking and bond formation during the processes but, at the same time, fast enough to sample a large number of configurations required to compute the corresponding free energy surfaces. Density functional theory (DFT) is often used to explore Pt-NPs; nonetheless, it is usually limited to some minimum-energy reaction pathways on static potential energy surfaces because of its high computational cost. In this work, we report a combination of the density functional tight binding (DFTB) method as a fast but reliable approximation to DFT, the steered molecular dynamics (SMD) technique, and the Jarzynski equality to construct free energy surfaces of the temperature-dependent diffusion and growth of platinum particles on a titania surface. In particular, we present the parametrization for Pt-X (X = Pt, Ti, or O) interactions in the framework of the second-order DFTB method, using a previous parametrization for titania as a basis. The optimized parameter set was used to simulate the surface diffusion of a single platinum atom (Pt 1 ) and the growth of Pt 6 from Pt 5 and Pt 1 on the rutile (110) surface at three different temperatures (T = 400, 600, 800 K). The free energy profile was constructed by using over a hundred SMD trajectories for each process. We found that increasing the temperature has a minimal effect on the formation free energy; nevertheless, it significantly reduces the free energy barrier of Pt atom migration on the TiO 2 surface and the transition state (TS) of its deposition. In a concluding remark, the methodology opens the pathway to quantum chemical free energy simulations of Pt-NPs’ temperature-dependent growth and other transformation processes on the titania support.
37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
Improving materials property predictions for graph neural networks with minimal feature engineering *
Graph neural networks (GNNs) have been employed in materials research to predict physical and functional properties, and have achieved superior performance in several application domains over prior machine learning approaches. Recent studies incorporate features of increasing complexity such as Gaussian radial functions, plane wave functions, and angular terms to augment the neural network models, with the expectation that these features are critical for achieving a high performance. Here, we propose a GNN that adopts edge convolution where hidden edge features evolve during training and extensive attention mechanisms, and operates on simple graphs with atoms as nodes and distances between them as edges. As a result, the same model can be used for very different tasks as no other domain-specific features are used. With a model that uses no feature engineering, we achieve performance comparable with state-of-the-art models with elaborate features for formation energy and band gap prediction with standard benchmarks; we achieve even better performance when the dataset size increases. Although some domain-specific datasets still require hand-crafted features to achieve state-of-the-art results, our selected architecture choices greatly reduce the need for elaborate feature engineering and still maintain predictive power in comparison.
42 ENGINEERING↗
Toward Understanding and Controlling Organic Reactions on Metal Oxide Catalysts
Metal oxides have structurally complex surfaces on which a variety of adsorption site types can occur, including cation sites, anion sites, oxygen vacancy sites, and Brønsted acid sites. These sites can catalyze the catalytic transformation of organic molecules via diverse routes, thus enabling H abstraction, O abstraction, C–C bond formation, and other reactions. This Perspective provides an update on recent advances and future directions for various organic reactions on metal oxide catalyst surfaces, particularly for C–H activation of alkanes and for C–C bond formation with organic oxygenate reactants. Here, we put emphasis on the molecular scale details, on the active site structures required to enable the formation of kinetically relevant transition states, energetic descriptors, as well as contemporary ideas to enable low activation energies. This progress has been enabled by specialized experiments and the increased capabilities of modern electronic structure calculations.
37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
In Situ Neutron Scattering Study of the Structure Dynamics of the Ru/Ca 2 N:e – Catalyst in Ammonia Synthesis
NH 3 synthesis is one of the most critical industrial processes. Compared to commercial iron catalysts, Ru catalysts show high intrinsic activity in this reaction but suffer from hydrogen poisoning. By loading Ru onto supports such as electrides and hydrides, the hydrogen poisoning problem can be significantly alleviated. However, relevant studies on the structural dynamics of the Ru/electride catalysts under reaction conditions are very scarce. Taking advantage of the high sensitivity to hydrogen species, it is possible to obtain insights into the structural changes during the reaction using in situ neutron techniques. In this study, we have investigated the structural evolution of the Ru/Ca 2 N:e – catalyst during the ammonia synthesis reaction by in situ neutron scattering (inelastic neutron scattering, INS) technique. In situ INS experiments suggest that Ca 2 N:e – is likely converted to the Ca 2 NH phase during the reaction. Unlike the previously known structure where H and N atoms are intermixed, the formed Ca 2 NH exhibits a segregated structure where the H and N atoms are located in different layers separated by the Ca layer. In conclusion, density functional theory calculations of the reaction energetics reveal that there are minor changes in the barriers and thermodynamics of the first N hydrogenation step between the two phases (Ca 2 NH phase with segregated H/N layers and intermixed Ca 2 NH phase), suggesting the impact of the phases on the reaction kinetics to be relatively minimal.
37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
Extensive Attention Mechanisms in Graph Neural Networks for Materials Discovery
We present our research where attention mechanism is extensively applied to various aspects of graph neural net- works for predicting materials properties. As a result, surrogate models can not only replace costly simulations for materials screening but also formulate hypotheses and insights to guide further design exploration. We predict formation energy of the Materials Project and gas adsorption of crystalline adsorbents, and demonstrate the superior performance of our graph neural networks. Moreover, attention reveals important substructures that the machine learning models deem important for a material to achieve desired target properties. Our model is based solely on standard structural input files containing atomistic descriptions of the adsorbent material candidates. We construct novel methodological extensions to match the prediction accuracy of state-of-the-art models some of which were built with hundreds of features at much higher computational cost. We show that sophisticated neural networks can obviate the need for elaborate feature engineering. Our approach can be more broadly applied to optimize gas capture processes at industrial scale.
Cong, Guojing↗
CH 4 Activation over Perovskite Catalysts: True Density and Reactivity of Active Sites
The high thermal stability of perovskites has drawn attention toward their applications for catalytic CH 4 activation and conversion, typically occurring at high temperatures. The reaction rates of perovskite catalysts for CH 4 combustion, however, trail behind those of noble metal catalysts. Ways to optimize the performance of perovskite catalysts are destined to trial-and-error approaches unless their complex reconstructed surfaces are correlated with fundamental kinetic parameters. Discerning the intrinsic activity of surface catalytic sites and the density of those sites is crucial to rationally envision complex metal oxides with enhanced catalytic performance. Here, the present work presents a detailed kinetic analysis of catalytic CH 4 combustion over a set of seven perovskites (SrTiO 3 , SrZrO 3 , SrFeO 3 , LaFeO 3 , LaInO 3 , LaCoO 3 , LaMnO 3 ) with various surface terminations. Steady-state isotopic transient kinetic analysis was employed to measure turnover frequency (TOF) and density of surface intermediates (N) under operando conditions. Top surface characterization elucidated performance-structure relationships between near-monolayer surface composition and intrinsic reactivity of the catalysts. By using a chemical etching procedure to expose Fe-sites at the top surface of LaFeO 3 (LaFeO 3 , HNO 3 ), its TOF was increased 4-fold, compared with the unmodified sample, although N on the surface of LaFeO 3 , HNO 3 decreased. Density functional theory simulations corroborated that surface Fe-termination and La-Fe termination offer lower energetic barriers for CH 4 activation when compared with La-termination. In general, surface reconstruction is shown as a tool to tune TOF and N to improve reaction rates. This work fills a gap in current kinetic studies of perovskites through a careful assessment and discussion of the density and intrinsic reactivity of active sites for methane combustion over well-characterized reconstructed perovskite surfaces.
37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
CO[subscript 2]-Assisted Oxidative Dehydrogenation of Propane over VO[subscript x]/In[subscript 2]O[
Explore the source record for details and available documents.
Jiang, Xiao↗
Measuring and directing charge transfer in heterogenous catalysts
Abstract Precise control of charge transfer between catalyst nanoparticles and supports presents a unique opportunity to enhance the stability, activity, and selectivity of heterogeneous catalysts. While charge transfer is tunable using the atomic structure and chemistry of the catalyst-support interface, direct experimental evidence is missing for three-dimensional catalyst nanoparticles, primarily due to the lack of a high-resolution method that can probe and correlate both the charge distribution and atomic structure of catalyst/support interfaces in these structures. We demonstrate a robust scanning transmission electron microscopy (STEM) method that simultaneously visualizes the atomic-scale structure and sub-nanometer-scale charge distribution in heterogeneous catalysts using a model Au-catalyst/SrTiO 3 -support system. Using this method, we further reveal the atomic-scale mechanisms responsible for the highly active perimeter sites and demonstrate that the charge transfer behavior can be readily controlled using post-synthesis treatments. This methodology provides a blueprint for better understanding the role of charge transfer in catalyst stability and performance and facilitates the future development of highly active advanced catalysts.
37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
Physically Informed Machine Learning Prediction of Electronic Density of States
The electronic structure of a material, such as its density of states (DOS), provides key insights into its physical and functional properties and serves as a valuable source of high-quality features for many materials screening and discovery workflows. Still, the computational cost of calculating the DOS, most commonly with density functional theory (DFT), becomes prohibitive for meeting high-fidelity or high-throughput requirements, necessitating a cheaper but sufficiently accurate surrogate. To fulfill this demand, we develop a general machine learning method based on graph neural networks for predicting the DOS purely from atomic positions, six orders of magnitude faster than DFT. This approach can effectively use large materials databases and be applied generally across the entire periodic table to materials classes of arbitrary compositional and structural diversity. We furthermore devise a highly adaptable scheme for physically informed learning which encourages the DOS prediction to favor physically reasonable solutions defined by any set of desired constraints. This functionality provides a means for ensuring that the predicted DOS is reliable enough to be used as an input to downstream materials screening workflows to predict more complex functional properties, which rely on accurate physical features.
36 MATERIALS SCIENCE↗
Revealing the interplay between “intelligent behavior” and surface reconstruction of non-precious metal doped SrTiO 3 catalysts during methane combustion
The impact of surface reconstruction of a model perovskite, SrTiO 3 (STO), on CH 4 activation for combustion and oxidative coupling was previously revealed that the reaction rate was proportional to the creation of Srterminated step sites. Doped perovskites (SrTi 1-x M x O 3 , M=metal dopant) present yet another form of reconstruction throughout the surface and the bulk, where the metal dopant can migrate in and out of the perovskite lattice, also known as "intelligent behavior". In this work, understanding the interplay between perovskite surface reconstruction (surface termination) and the "intelligent behavior" is tackled for the first time, and the catalytic consequences are probed with CH 4 combustion as a model reaction. A set of experimental techniques, including XRD, Raman spectroscopy, X-ray adsorption spectroscopy, kinetic measurements, as well as DFT calculations were used to understand the catalytic behavior of the reconstructed surfaces of Ni and Cu-doped STO for methane combustion. Here, we found that during methane oxidation, the diffusion of Ni and Cu into the lattice due to the "intelligent behavior" is accompanied by Sr enrichment on the surface of the perovskite. This Srenrichment process is reversible when Cu or Ni species exsolute as clusters/nanoparticles upon H 2 treatment. Such a surface reconstruction is found to greatly impact the catalytic activity of doped perovskites towards methane combustion.
37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
Inverse design of two-dimensional materials with invertible neural networks
The ability to readily design novel materials with chosen functional properties on-demand represents a next frontier in materials discovery. However, thoroughly and efficiently sampling the entire design space in a computationally tractable manner remains a highly challenging task. To tackle this problem, we propose an inverse design framework (MatDesINNe) utilizing invertible neural networks which can map both forward and reverse processes between the design space and target property. This approach can be used to generate materials candidates for a designated property, thereby satisfying the highly sought-after goal of inverse design. We then apply this framework to the task of band gap engineering in two-dimensional materials, starting with MoS 2 . Within the design space encompassing six degrees of freedom in applied tensile, compressive and shear strain plus an external electric field, we show the framework can generate novel, high fidelity, and diverse candidates with near-chemical accuracy. We extend this generative capability further to provide insights regarding metal-insulator transition in MoS 2 which are important for memristive neuromorphic applications, among others. This approach is general and can be directly extended to other materials and their corresponding design spaces and target properties.
36 MATERIALS SCIENCE↗
Benchmarking graph neural networks for materials chemistry
Graph neural networks (GNNs) have received intense interest as a rapidly expanding class of machine learning models remarkably well-suited for materials applications. To date, a number of successful GNNs have been proposed and demonstrated for systems ranging from crystal stability to electronic property prediction and to surface chemistry and heterogeneous catalysis. However, a consistent benchmark of these models remains lacking, hindering the development and consistent evaluation of new models in the materials field. Here, we present a workflow and testing platform, MatDeepLearn, for quickly and reproducibly assessing and comparing GNNs and other machine learning models. We use this platform to optimize and evaluate a selection of top performing GNNs on several representative datasets in computational materials chemistry. From our investigations we note the importance of hyperparameter selection and find roughly similar performances for the top models once optimized. We identify several strengths in GNNs over conventional models in cases with compositionally diverse datasets and in its overall flexibility with respect to inputs, due to learned rather than defined representations. Meanwhile several weaknesses of GNNs are also observed including high data requirements, and suggestions for further improvement for applications in materials chemistry are discussed.
36 MATERIALS SCIENCE↗
New Insights into the Bulk and Surface Defect Structures of Ceria Nanocrystals from Neutron Scattering Study
Neutron diffraction and pair distribution function studies coupled with Raman spectroscopy have successfully unraveled the detailed oxygen defect structures of ceria nanocubes and nanorods. Two types of defect sites are revealed for the ceria nanocrystals: surface and bulk defects. It is proposed that the surface oxygen defects in both types of CeO 2 nanocrystals are predominantly the partially reduced Ce 3 O 5+x , with the bulk defect structures dominated by interstitial Frenkel-type oxygen vacancies. Ceria nanorods possess much higher concentration of surface oxygen defects relative to the nanocubes, albeit with only slightly higher concentration of bulk Frenkel-type oxygen vacancies. Upon annealing the nanorod sample at 600 °C under vacuum (~10 –4 to 10 –5 mbar), a partially reduced ceria phase with long-range oxygen vacancy ordering (Ce 3 O 5+x ) has been observed experimentally for the first time. This intriguing observation that surface defect phases can take on ordered defect sublattices under certain conditions is of great value in understanding the temperature-dependent catalytic performance of ceria nanocrystals. Furthermore, a drastic decrease of the surface vacancies in the ceria nanocrystals is observed upon exposure to SO 2 , especially for the nanorods, a likely origin for the sulfur poisoning effect on ceria-based materials. This study suggests that tailoring surface morphology is a promising strategy to control defect properties of ceria nanomaterials. It also provides fundamental insights to stabilize surface oxygen defects in CeO 2 nanocrystals to achieve high redox performance under corrosive environments such as under SO 2 /SO x exposure.
36 MATERIALS SCIENCE↗
Towards Efficient Uncertainty estimation in deep learning for robust energy prediction in crystal materials
In material science, recent studies have started to explore the potential of using deep learning to improve property prediction from high-fidelity simulations, e.g, density functional theory (DFT). However, the design spaces are sometimes too large and intractable to sample completely. This results in a critical question that is how to evaluate the confidence and robustness of the prediction. In this paper, we propose an efficient approach to estimate uncertainty in deep learning using a single forward pass and then apply it for robust prediction of the total energy in crystal lattice structures. Our approach is built upon the deep kernel learning (DKL) that originally introduces to leverage the expressiveness of deep neural networks as input with a probabilistic prediction of Gaussian processes (GPs) as output. Existing DKL methods have difficulties in the accuracy of predictive uncertainty, training stability, and scaling to large datasets, which lead to significant barriers in real-world applications. We propose to address these challenges by using an inducing point approximate GP in feature space combined with spectral normalization as a regularization. We finally demonstrate our robust performance on an artificial example and a real-world application from materials chemistry.
Bi, Sirui↗
Work Function Engineering of 2D Materials: The Role of Polar Edge Reconstructions
2D materials have attracted tremendous interest as functional materials because of their diverse and tunable properties, especially at their edges. A material’s work function is a critical parameter in many electronic devices; however, a fundamental understanding and a path toward large alterations of the work function in 2D materials still remain elusive. Here in this paper, we report the first evidence for anisotropy of the work function in 2D MoS 2 from first-principles calculations. We also demonstrate large work-function tunability (in the range of 3.45–6.29 eV) choosing the 2H phase of MoS 2 as a model system by sampling various edge configurations. We furthermore reveal the origin of this work function anisotropy and tunability by extending the existing work function relation to the local dipole moment at surfaces of 3D materials to those at edges in 2D materials. We then use machine-learning approaches to correlate work function with edge structures. These results pave the way for intrinsic edge engineering for electronic and catalytic applications.
36 MATERIALS SCIENCE↗
Oxidative Dehydrogenation of Propane to Propylene with Soft Oxidants via Heterogeneous Catalysis
Oxidative dehydrogenation of propane to propylene can be achieved using conventional, oxygen-assisted dehydrogenation of propane (O 2 –ODHP) or via the use of soft oxidants, such as CO 2 , N 2 O, S-containing compounds, and halogens/halides. The major roles of soft oxidants include inhibiting overoxidation and improving propylene selectivity, which are considered to be current challenges in O 2 -assisted dehydrogenation. For both CO 2 – and N 2 O–ODHP reactions, significant efforts have been devoted to developing redox-active (e.g., chromium, vanadate, iron, etc.), nonredox-type main group metal oxide (e.g., group IIIA, gallium), and other transition metal/metal oxide catalysts (e.g., molybdenum, palladium platinum, rhodium, ruthenium, etc.), as well as zeolite-based catalysts with adjustable acid–base properties, unique pore structures, and topologies. Metal sulfides have shown promising performance in DHP, whereas the development of suitable catalysts has lagged for SO 2 - or S-assisted ODHP. Recently, significant efforts have been focused on homogeneous and heterogeneous ODHP using halogens (e.g., Br 2 , I 2 , Cl 2 , etc.) and hydrogen halides (e.g., HCl and HBr) for the development of facile processes for C 3 H 6 synthesis. This work aims to provide a critical, comprehensive review of recent advances in oxidative dehydrogenation of propane with these soft oxidants, especially highlighting the current state of understanding of the following factors: (i) relationships between composition, structure, and catalytic performance, (ii) effects of the support, acidity, and promoters, (iii) reaction pathway and mechanistic insights, and (iv) the various roles of soft oxidants. Theoretical and computational insights toward understanding reaction mechanisms and catalyst design principles are also covered. Future research opportunities are discussed in terms of catalyst design and synthesis, deactivation and regeneration, reaction mechanisms, and alternative approaches.
37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗