arXiv subjects

Nathan Keilbart

Publications and source records attributed to Nathan Keilbart.

5 recordsLinked to original sources

arXiv2025

Active Learning Enables Extrapolation in Molecular Generative Models

Although generative models hold promise for discovering molecules with optimized desired properties, they often fail to suggest synthesizable molecules that improve upon the known molecules seen in training. We find that a key limitation is not in the molecule generation process itself, but in the poor generalization capabilities of molecular property predictors. We tackle this challenge by creating an active-learning, closed-loop molecule generation pipeline, whereby molecular generative models are iteratively refined on feedback from quantum chemical simulations to improve generalization to new chemical space. Compared against other generative model approaches, only our active learning approach generates molecules with properties that extrapolate beyond the training data (reaching up to 0.44 standard deviations beyond the training data range) and out-of-distribution molecule classification accuracy is improved by 79%. By conditioning molecular generation on thermodynamic stability data from the active-learning loop, the proportion of stable molecules generated is 3.5x higher than the next-best model.

cs.LG↗

arXiv2022

Hydrogen in Disordered Titania: Connecting Local Chemistry, Structure, and Stoichiometry through Accelerated Exploration

Hydrogen incorporation in native surface oxides of metal alloys often controls the onset of metal hydriding, with implications for materials corrosion and hydrogen storage. A key representative example is titania, which forms as a passivating layer on a variety of titanium alloys for structural and functional applications. These oxides tend to be structurally diverse, featuring polymorphic phases, grain boundaries, and amorphous regions that generate a disparate set of unique local environments for hydrogen. Here, we introduce a workflow that can efficiently and accurately navigate this complexity. First, a machine learning force field, trained on ab initio molecular dynamics simulations, was used to generate amorphous configurations. Density functional theory calculations were then performed on these structures to identify local oxygen environments, which were compared against experimental observations. Second, to classify subtle differences across the disordered configuration space, we employ a graph-based sampling procedure. Finally, local hydrogen binding energies are computed using exhaustive density functional theory calculations on representative configurations. We leverage this methodology to show that hydrogen binding energetics are described by local oxygen coordination, which in turn is affected by stoichiometry. Together these results imply that hydrogen incorporation and transport in TiO$_x$ can be tailored through compositional engineering, with implications for improving performance and durability of titanium-derived alloys in hydrogen environments.

cond-mat.mtrl-sci↗

arXiv2021

Efficient, Interpretable Graph Neural Network Representation for Angle-dependent Properties and its Application to Optical Spectroscopy

Graph neural networks are attractive for learning properties of atomic structures thanks to their intuitive graph encoding of atoms and bonds. However, conventional encoding does not include angular information, which is critical for describing atomic arrangements in disordered systems. In this work, we extend the recently proposed ALIGNN encoding, which incorporates bond angles, to also include dihedral angles (ALIGNN-d). This simple extension leads to a memory-efficient graph representation that captures the complete geometry of atomic structures. ALIGNN-d is applied to predict the infrared optical response of dynamically disordered Cu(II) aqua complexes, leveraging the intrinsic interpretability to elucidate the relative contributions of individual structural components. Bond and dihedral angles are found to be critical contributors to the fine structure of the absorption response, with distortions representing transitions between more common geometries exhibiting the strongest absorption intensity. Future directions for further development of ALIGNN-d are discussed.

cs.LG↗

arXiv2021

Using DFTB to Model Photocatalytic Anatase-Rutile TiO$_2$ Nanocrystalline Interfaces and their Band Alignment

Band alignment effects of anatase and rutile nanocrystals in TiO$_2$ powders lead to an electron hole separation, increasing the photo catalytic efficiency of these powders. While size effects and types of possible alignments have been extensively studied, the effect of interface geometries of bonded nanocrystal structures on the alignment is poorly understood. In order to allow conclusive studies of a vast variety of bonded systems in different orientations, we have developed a new density functional tight binding parameter set to properly describe quantum confinement in nanocrystals. Applying this set we found a quantitative influence of the interface structure on the band alignment.

cond-mat.mtrl-sci↗

arXiv2017

Quantum-continuum simulation of the electrochemical response of pseudocapacitor electrodes from realistic conditions

Pseudocapacitors are energy-storage devices characterized by fast and reversible redox reactions that enable them to store large amounts of electrical energy at high rates. We simulate the response of pseudocapacitive electrodes under realistic conditions to identify the microscopic factors that determine their performance, focusing on ruthenia (RuO2) as a prototypical electrode material. Electronic-structure methods are used together with a self-consistent continuum solvation (SCCS) model to build a complete dataset of free energies as the surface of the charged electrode is gradually covered with protons under applied voltage. The resulting dataset is exploited to compute hydrogen-adsorption isotherms and charge-voltage responses by means of grand-canonical sampling, finding close agreement with experimental voltammetry. These simulations reveal that small changes on the order of 5 μF/cm2 in the intrinsic double-layer capacitance of the electrode-electrolyte interface can induce variations of up to 40 μF/cm2 in the overall pseudocapacitance.

cond-mat.mtrl-sci↗