Search

We demonstrate the ability of large language models (LLMs) to perform material and molecular property regression tasks, a significant deviation from the conventional LLM use case. We benchmark the Large Language Model Meta AI (LLaMA) 3 on several molecular properties in the QM9 dataset and 24 materials properties. Only composition-based input strings are used as the model input and we fine tune on only the generative loss. We broadly find that LLaMA 3, when fine-tuned using the SMILES representation of molecules, provides useful regression results which can rival standard materials property prediction models like random forest or fully connected neural networks on the QM9 dataset. Not surprisingly, LLaMA 3 errors are 5-10x higher than those of the state-of-the-art models that were trained using far more granular representation of molecules (e.g., atom types and their coordinates) for the same task. Interestingly, LLaMA 3 provides improved predictions compared to GPT-3.5 and GPT-4o. This work highlights the versatility of LLMs, suggesting that LLM-like generative models can potentially transcend their traditional applications to tackle complex physical phenomena, thus paving the way for future research and applications in chemistry, materials science and other scientific domains.

The integration of machine learning (ML) models enhances the efficiency, affordability, and reliability of feature detection in microscopy, yet their development and applicability are hindered by the dependency on scarce and often flawed manually labeled datasets and a lack of domain awareness. We addressed these challenges by creating a physics-based synthetic image and data generator, resulting in a machine learning model that achieves comparable precision (0.86), recall (0.63), F1 scores (0.71), and engineering property predictions (R2=0.82) to a model trained on human-labeled data. We enhanced both models by using feature prediction confidence scores to derive an image-wide confidence metric, enabling simple thresholding to eliminate ambiguous and out-of-domain images resulting in performance boosts of 5-30% with a filtering-out rate of 25%. Our study demonstrates that synthetic data can eliminate human reliance in ML and provides a means for domain awareness in cases where many feature detections per image are needed.

One compelling vision of the future of materials discovery and design involves the use of machine learning (ML) models to predict materials properties and then rapidly find materials tailored for specific applications. However, realizing this vision requires both providing detailed uncertainty quantification (model prediction errors and domain of applicability) and making models readily usable. At present, it is common practice in the community to assess ML model performance only in terms of prediction accuracy (e.g., mean absolute error), while neglecting detailed uncertainty quantification and robust model accessibility and usability. Here, we demonstrate a practical method for realizing both uncertainty and accessibility features with a large set of models. We develop random forest ML models for 33 materials properties spanning an array of data sources (computational and experimental) and property types (electrical, mechanical, thermodynamic, etc.). All models have calibrated ensemble error bars to quantify prediction uncertainty and domain of applicability guidance enabled by kernel-density-estimate-based feature distance measures. All data and models are publicly hosted on the Garden-AI infrastructure, which provides an easy-to-use, persistent interface for model dissemination that permits models to be invoked with only a few lines of Python code. We demonstrate the power of this approach by using our models to conduct a fully ML-based materials discovery exercise to search for new stable, highly active perovskite oxide catalyst materials.

Oxygen ion conductors are crucial for enhancing the efficiency of various clean energy technologies, including fuel cells, batteries, electrolyzers, membranes, sensors, and more. In this study, LaBi2O4Cl is identified as an ultra-fast oxygen conductor from the MBi2O4X (M=rare-earth element, X=halogen element) family, discovered by a structure-similarity analysis of >60k oxygen-containing compounds. Ab initio studies reveal that LaBi2O4Cl has an ultra-low migration barrier of 0.1 eV for oxygen vacancy, significantly lower than 0.6-0.8 eV for interstitial oxygen. Frenkel pairs are the dominant defects in intrinsic LaBi2O4Cl, facilitating notable oxygen diffusion primarily through vacancies at higher temperatures. LaBi2O4Cl with extrinsic oxygen vacancies (2.8%) exhibits a conductivity of 0.3 S/cm at 25{\deg}C, maintains a 0.1 eV diffusion barrier up to 1100{\deg}C, and transitions from extrinsic to mixed extrinsic and intrinsic behavior as the Frenkel pair concentration increases at higher temperatures. Experimental results on synthesized LaBi2O4Cl and Sr-doped LaBi2O4Cl demonstrate comparable or higher oxygen conductivity than YSZ and LSGM below 400 {\deg}C, with lower activation energies. Further experimental optimization of LaBi2O4Cl, including aliovalent doping and microstructure refinement, could significantly enhance its performance and efficiency, facilitating fast oxygen conduction approaching room temperature.

Knowledge of the domain of applicability of a machine learning model is essential to ensuring accurate and reliable model predictions. In this work, we develop a new approach of assessing model domain and demonstrate that our approach provides accurate and meaningful designation of in-domain versus out-of-domain when applied across multiple model types and material property data sets. Our approach assesses the distance between a test and training data point in feature space by using kernel density estimation and shows that this distance provides an effective tool for domain determination. We show that chemical groups considered unrelated based on established chemical knowledge exhibit significant dissimilarities by our measure. We also show that high measures of dissimilarity are associated with poor model performance (i.e., high residual magnitudes) and poor estimates of model uncertainty (i.e., unreliable uncertainty estimation). Automated tools are provided to enable researchers to establish acceptable dissimilarity thresholds to identify whether new predictions of their own machine learning models are in-domain versus out-of-domain., Comment: 43 pages, 34 figures, journal submission

Perovskite SrVO$_3$ has recently been proposed as a novel electron emission cathode material. Density functional theory (DFT) calculations suggest multiple low work function surfaces and recent experimental efforts have consistently demonstrated effective work functions of ~2.7 eV for polycrystalline samples, both results suggesting, but not directly confirming, some fraction of even lower work function surface is present. In this work, thermionic electron emission microscopy (ThEEM) and high-field ultraviolet photoemission spectroscopy are used to study the local work function distribution and measure the work function of a partially-oriented-(110)-SrVO$_3$ perovskite oxide cathode surface. Our results show direct evidence of low work function patches of about 2.1 eV on the cathode surface, with corresponding onset of observable thermionic emission at 750 $^o$C. We hypothesize that, in our ThEEM experiments, the high applied electric field suppresses the patch field effect, enabling the direct measurement of local work functions. This measured work function of 2.1 eV is comparable to the previous DFT-calculated work function value of the SrVO-terminated (110) SrVO$_3$ surface (2.3 eV) and SrO terminated (100) surface (1.9 eV). The measured 2.1 eV value is also much lower than the work function for the (001) LaB$_6$ single crystal cathode (~2.7 eV) and comparable to the effective work function of B-type dispenser cathodes (~2.1 eV). If SrVO$_3$ thermionic emitters can be engineered to access domains of this low 2.1 eV work function, they have potential to significantly improve thermionic emitter-based technologies.

Discovering new materials that efficiently catalyze the oxygen reduction and evolution reactions is critical for facilitating the widespread adoption of solid oxide fuel cell and electrolyzer (SOFC/SOEC) technologies. Here, we develop machine learning (ML) models to predict perovskite catalytic properties critical for SOFC/SOEC applications, including oxygen surface exchange, oxygen diffusivity, and area specific resistance (ASR). The models are based on trivial-to-calculate elemental features and are more accurate and dramatically faster than the best models based on ab initio-derived features, potentially eliminating the need for ab initio calculations in descriptor-based screening. Our model of ASR enables temperature-dependent predictions, has well calibrated uncertainty estimates and online accessibility. Use of temporal cross-validation reveals our model to be effective at discovering new promising materials prior to their initial discovery, demonstrating our model can make meaningful predictions. Using the SHapley Additive ExPlanations (SHAP) approach, we provide detailed discussion of different approaches of model featurization for ML property prediction. Finally, we use our model to screen more than 19 million perovskites to develop a list of promising cheap, earth-abundant, stable, and high performing materials, and find some top materials contain mixtures of less-explored elements (e.g., K, Bi, Y, Ni, Cu) worth exploring in more detail.

The discovery and design of new materials which can efficiently catalyze the oxygen reduction and evolution reactions at reduced temperatures is important for facilitating the widespread adoption of fuel cell and electrolyzer technologies. Numerous studies have produced correlations between catalytic properties, such as oxygen surface exchange or electrode area specific resistance (ASR), and properties of the catalyst material. However, correlations have historically been limited in scope (e.g., using only a few materials or at a single temperature) and it has been difficult to provide detailed assessments of their robustness. Here, we assess the ability of the O p-band center electronic structure descriptor, obtained from density functional theory (DFT) calculations, to correlate with oxygen surface exchange rates, diffusivities, and area specific resistances for a large database of perovskite oxide catalytic properties. By data mining the literature, we obtain 747 catalytic property value data points spanning 299 unique perovskite compositions from 313 studies. We assess linear correlations of each property with the O p-band center and find generally modest correlations that are qualitatively useful (prediction mean absolute errors of about 0.5 log units are typical), where the correlations are improved at higher temperatures (e.g., 800 {\deg}C vs. 500 {\deg}C) and significantly improve when considering fits to the subset of materials which have multiple independent measurements. These findings suggest that the spread of property data is significantly influenced by experimental uncertainty, and subsequent measurements of additional materials will likely improve the O p-band center correlations.

Thermionic electron emission cathodes are critical components of various high power and high frequency vacuum electronic devices, electron microscopes, e-beam lithographic devices, and thermionic energy converters, which all demand an efficient and long-lasting low work function cathode. Single phase, polycrystalline perovskite oxide SrVO$_3$, with its intrinsic low effective work function and facile synthesis process, is a promising cathode candidate, where previous works have shown evidence of an effective work function as low as 2.3 eV. However, assessment of the stability over time under conditions relevant for operation and the related interplay of evolving surface chemistry with emission performance are still missing, and necessary for understanding how to best prepare, process and operate SrVO$_3$ cathodes. In this work, we study the vacuum activation process of SrVO$_3$ and find it has promising emission stability over 15 days of continuous high temperature operation. We find that SrVO$_3$ shows surface Sr and O segregation during operation, which we hypothesize is needed to create a positive surface dipole, leading to low effective work function. Emission repeatability from cyclic heating and cooling suggests the promising stability of the low effective work function surface, and additional observations of drift-free emission during one hour of continuous emission testing at high temperature further demonstrates its excellent performance stability.

Materials discovery and design typically proceeds through iterative evaluation (both experimental and computational) to obtain data, generally targeting improvement of one or more properties under one or more constraints (e.g., time or budget). However, there can be great variation in the quality and cost of different data, and when they are mixed together in what we here call multifidelity data the optimal approaches to their utilization are not established. It is therefore important to develop strategies to acquire and use multifidelity data to realize the most efficient iterative materials exploration. In this work, we assess the impact of using multifidelity data through mock demonstration of designing solar cell materials, using the electronic bandgap as the target property. We propose a new approach of using multifidelity data through leveraging machine learning models of both low- and high-fidelity data, where using predicted low-fidelity data as an input feature in the high-fidelity model can improve the impact of a multifidelity data approach. We show how tradeoffs of low- versus high-fidelity measurement cost and acquisition can impact the materials discovery process, and find that the use of multifidelity data has maximal impact on the materials discovery campaign when approximately five low-fidelity measurements per high-fidelity measurement are performed, and when the cost of low-fidelity measurements is approximately 5% or less than that of high-fidelity measurements. This work provides practical guidance and useful qualitative measures for improving materials discovery campaigns that involve multifidelity data.

An essential aspect of extending safe operation of the active nuclear reactors is understanding and predicting the embrittlement that occurs in the steels that make up the Reactor pressure vessel (RPV). In this work we integrate state of the art machine learning methods using ensembles of neural networks with unprecedented data collection and integration to develop a new model for RPV steel embrittlement. The new model has multiple improvements over previous machine learning and hand-tuned efforts, including greater accuracy (e.g., at high-fluence relevant for extending the life of present reactors), wider domain of applicability (e.g., including a wide-range of compositions), uncertainty quantification, and online accessibility for easy use by the community. These improvements provide a model with significant new capabilities, including the ability to easily and accurately explore compositions, flux, and fluence effects on RPV steel embrittlement for the first time. Furthermore, our detailed comparisons show our approach improves on the leading American Society for Testing and Materials (ASTM) E900-15 standard model for RPV embrittlement on every metric we assessed, demonstrating the efficacy of machine learning approaches for this type of highly demanding materials property prediction.

New highly oxygen-active materials may enhance many energy-related technologies by enabling efficient oxygen-ion transport at lower temperatures, e.g., below 400 Celsius. Interstitial oxygen conductors have the potential to realize such performance but have received far less attention than vacancy-mediated conductors. Here, we combine physically-motivated structure and property descriptors, ab initio simulations, and experiments to demonstrate an approach to discover new fast interstitial oxygen conductors. Multiple new families were found which adopt completely different structures from known oxygen conductors. From these families, we synthesized and studied oxygen kinetics in La4Mn5Si4O22+d (LMS), a representative member of perrierite/chevkinite family. We found LMS has higher oxygen ionic conductivity than the widely used yttria-stabilized ZrO2, and among the highest surface oxygen exchange rates at intermediate temperature of known materials. The fast oxygen kinetics is the result of simultaneously active interstitial and interstitialcy diffusion pathways. This work developed and demonstrated a powerful approach for discovering new families of interstitial oxygen conductors and suggests many more such materials remain to be discovered.

Accurately quantifying swelling of alloys that have undergone irradiation is essential for understanding alloy performance in a nuclear reactor and critical for the safe and reliable operation of reactor facilities. However, typical practice is for radiation-induced defects in electron microscopy images of alloys to be manually quantified by domain-expert researchers. Here, we employ an end-to-end deep learning approach using the Mask Regional Convolutional Neural Network (Mask R-CNN) model to detect and quantify nanoscale cavities in irradiated alloys. We have assembled the largest database of labeled cavity images to date, which includes 400 images, >34k discrete cavities, and numerous alloy compositions and irradiation conditions. We have evaluated both statistical (precision, recall, and F1 scores) and materials property-centric (cavity size, density, and swelling) metrics of model performance, and performed in-depth analysis of materials swelling assessments. We find our model gives assessments of material swelling with an average (standard deviation) swelling mean absolute error based on random leave-out cross-validation of 0.30 (0.03) percent swelling. This result demonstrates our approach can accurately provide swelling metrics on a per-image and per-condition basis, which can provide helpful insight into material design (e.g., alloy refinement) and impact of service conditions (e.g., temperature, irradiation dose) on swelling. Finally, we find there are cases of test images with poor statistical metrics, but small errors in swelling, pointing to the need for moving beyond traditional classification-based metrics to evaluate object detection models in the context of materials domain applications.

Quantifying charge-state transition energy levels of impurities in semiconductors is critical to understanding and engineering their optoelectronic properties for applications ranging from solar photovoltaics to infrared lasers. While these transition levels can be measured and calculated accurately, such efforts are time-consuming and more rapid prediction methods would be beneficial. Here, we significantly reduce the time typically required to predict impurity transition levels using multi-fidelity datasets and a machine learning approach employing features based on elemental properties and impurity positions. We use transition levels obtained from low-fidelity (i.e., local-density approximation or generalized gradient approximation) density functional theory (DFT) calculations, corrected using a recently proposed modified band alignment scheme, which well-approximates transition levels from high-fidelity DFT (i.e., hybrid HSE06). The model fit to the large multi-fidelity database shows improved accuracy compared to the models trained on the more limited high-fidelity values. Crucially, in our approach, when using the multi-fidelity data, high-fidelity values are not required for model training, significantly reducing the computational cost required for training the model. Our machine learning model of transition levels has a root mean squared (mean absolute) error of 0.36 (0.27) eV vs high-fidelity hybrid functional values when averaged over 14 semiconductor systems from the II-VI and III-V families. As a guide for use on other systems, we assessed the model on simulated data to show the expected accuracy level as a function of bandgap for new materials of interest. Finally, we use the model to predict a complete space of impurity charge-state transition levels in all zinc blende III-V and II-VI systems.

In a current density versus temperature (J-T) (Miram) curve in thermionic electron emission, experimental measurements demonstrate there is a smooth transition between the exponential region and the saturated emission regions, which is sometimes referred to as the "roll-off" or "Miram curve knee". The shape of the Miram curve knee is an important figure of merit for thermionic vacuum cathodes. Specifically, cathodes with a sharp Miram curve knee at low temperature with a flat saturated emission current are typically preferred. Our previous work on modeling nonuniform thermionic emission revealed that the space charge effect and patch field effect are key pieces of physics which impact the shape of the Miram curve knee. This work provides a more complete understanding of the physical factors connecting these physical effects and their relative impact on the shape of the knee, including the smoothness, the temperature, and the flatness of the saturated emission current density. For our analyses, we use a periodic, equal-width striped ("zebra crossing") work function distribution as a model system and illustrate how the space charge and patch field effects restrict the emission current density near the Miram curve knee. The results indicate there are three main physical parameters which significantly impact the shape of the Miram curve. Such physical knowledge directly connects the patch size, work function values, anode-cathode voltage, and anode-cathode gap distance to the shape of the Miram curve, providing new understanding and a guide to the design of thermionic cathodes used as electron sources in vacuum electronic devices (VEDs).

This paper provides an accurate theoretical defect energy database for pure and Bi-containing III-V (III-V:Bi) materials and investigates efficient methods for high-throughput defect calculations based on corrections of results obtained with local and semi-local functionals. Point defects as well as nearest-neighbor and second-nearest-neighbor pair defects were investigated in charge states ranging from -5 to 5. Ga-V:Bi systems (GaP:Bi, GaAs:Bi, and GaSb:Bi) were thoroughly investigated with significantly slower, higher fidelity hybrid Heyd-Scuseria-Ernzerhof (HSE) and significantly faster, lower fidelity local density approximation (LDA) calculations. In both approaches spurious electrostatic interactions were corrected with the Freysoldt correction. The results were verified against available experimental results and used to assess the accuracy of a previous band alignment correction. Here, a modified band alignment method is proposed in order to better predict the HSE values from the LDA ones. The proposed method allows prediction of defect energies with values that approximate those from the HSE functional at the computational cost of LDA (about 20x faster for the systems studied here). Tests of selected point defects in In-V:Bi materials resulted in corrected LDA values having a mean absolute error (MAE)=0.175 eV for defect levels vs. HSE. The method was further verified on an external database of defects and impurities in CdX (X=S, Se, Te) systems, yielding a MAE=0.194 eV. These tests demonstrate the correction to be sufficient for qualitative and semi-quantitative predictions, and may suggest transferability to many semiconductor systems without significant loss in accuracy. Properties of the remaining In-V:Bi defects and all Al-V:Bi defects were predicted with the use of the modified band alignment method.

Experimental observations of thermionic electron emission demonstrate a smooth transition between TL and FSCL regions of the emitted-current-density-versus-temperature (J-T) (Miram) curve and the emitted-current-density-versus-voltage (J-V) curve. Knowledge of the temperature and shape of the TL-FSCL transition is important in evaluating the thermionic electron emission performance of cathodes, including predicting the lifetime. However, there have been no first-principles physics-based models that can predict the smooth TL-FSCL transition region for real thermionic cathodes without applying physically difficult to justify a priori assumptions or empirical phenomenological equations. Previous work detailing the nonuniform thermionic emission found that the effects of 3-D space charge, patch fields, and Schottky barrier lowering can lead to a smooth TL-FSCL transition region from a model thermionic cathode surface with a checkerboard spatial distribution of work function values. In this work, we construct a physics-based nonuniform emission model for commercial dispenser cathodes for the first time. This emission model is obtained by incorporating the cathode surface grain orientation via electron backscatter diffraction (EBSD) and the facet-orientation-specific work function values from density functional theory (DFT) calculations. The model enables construction of two-dimensional emitted current density maps of the cathode surface and corresponding J-T and J-V curves. The predicted emission curves show excellent agreement with experiment, not only in TL and FSCL regions but, crucially, also in the TL-FSCL transition region. This model improves the understanding on the relationship between thermionic emission and cathode microstructure, which is beneficial to the design of vacuum electronic devices.

In this work, we perform semantic segmentation of multiple defect types in electron microscopy images of irradiated FeCrAl alloys using a deep learning Mask Regional Convolutional Neural Network (Mask R-CNN) model. We conduct an in-depth analysis of key model performance statistics, with a focus on quantities such as predicted distributions of defect shapes, defect sizes, and defect areal densities relevant to informing modeling and understanding of irradiated Fe-based materials properties. To better understand the performance and present limitations of the model, we provide examples of useful evaluation tests which include a suite of random splits, and dataset size-dependent and domain-targeted cross validation tests. Overall, we find that the current model is a fast, effective tool for automatically characterizing and quantifying multiple defect types in microscopy images, with a level of accuracy on par with human domain expert labelers. More specifically, the model can achieve average defect identification F1 scores as high as 0.8, and, based on random cross validation, have low overall average (+/- standard deviation) defect size and density percentage errors of 7.3 (+/- 3.8)% and 12.7 (+/- 5.3)%, respectively. Further, our model predicts the expected material hardening to within 10-20 MPa (about 10% of total hardening), which is about the same error level as experiments. Our targeted evaluation tests also suggest the best path toward improving future models is not expanding existing databases with more labeled images but instead data additions that target weak points of the model domain, such as images from different microscopes, imaging conditions, irradiation environments, and alloy types. Finally, we discuss the first phase of an effort to provide an easy-to-use, open-source object detection tool to the broader community for identifying defects in new images.