Radiography is an imaging technique used in a variety of applications, such as medical diagnosis, airport security, and nondestructive testing. We present a deep learning system for extracting information from radiographic images. We perform various prediction tasks using our system, including material classification and regression on the dimensions of a given object that is being radiographed. Our system is designed to address the sparse-data issue for radiographic nondestructive testing applications. It uses a radiographic simulation tool for synthetic data augmentation, and it uses transfer learning with a pre-trained convolutional neural network model. Using this system, our preliminary results indicate that the object geometry regression task saw an improvement of 70% in the R-squared value when using a multi-regime model. In addition, we increase the performance of the object material classification tasks by utilizing data from different imaging systems. In particular, using neutron imaging improved the material classification accuracy by 20% when compared to x-ray imaging.

Traditionally, developing new language models (LMs) capable of addressing multiple tasks involves fine-tuning pre-trained LMs using a wide collection of datasets, a process that often incurs significant computational expenses. Model merging emerges as a cost-effective alternative, allowing the integration of existing models fine-tuned on different tasks into a single model that performs well across all tasks, eliminating the need for additional training. In this paper, we propose RankMean, an algorithm for merging fine-tuned LMs without requiring any downstream data. RankMean determines merging coefficients based on the relative rankings of weight change magnitudes and applies these coefficients for module-wise integration of various fine-tuned models. Our experimental results demonstrate that RankMean outperforms existing baseline methods on multiple benchmarks. The code is available at github.com/VITA-Group/RankMean.

Predicting phase stability in high entropy alloys (HEAs), such as phase fractions as functions of composition and temperature, is essential for understanding alloy properties and screening desirable materials. Traditional methods like CALPHAD are computationally intensive for exploring high-dimensional compositional spaces. To address such a challenge, this study explored and compared the effectiveness of random forests (RF) and deep neural networks (DNN) for accelerating materials discovery by building surrogate models of phase stability prediction. For interpolation scenarios (testing on the same order of system as trained), RF models generally produce smaller errors than DNN models. However, for extrapolation scenarios (training on lower-order systems and testing on higher order systems), DNNs generalize more effectively than traditional ML models. DNN demonstrate the potential to predict topologically relevant phase composition when data were missing, making it a powerful predictive tool in materials discovery frameworks. The study uses a CALPHAD dataset of 480 million data points generated from a custom database, available for further model development and benchmarking. Experiments show that DNN models are data-efficient, achieving similar performance with a fraction of the dataset. This work highlights the potential of DNNs in materials discovery, providing a powerful tool for predicting phase stability in HEAs, particularly within the Cr-Hf-Mo-Nb-Ta-Ti-V-W-Zr composition space.

Discovering high-entropy alloys (HEAs) with high yield strength is an important yet challenging task in material science. However, the yield strength can only be accurately measured by very expensive and time-consuming real-world experiments, hence cannot be acquired at scale. Learning-based methods could facilitate the discovery process, but the lack of a comprehensive dataset on HEA yield strength has created barriers. We present X-Yield, a large-scale material science benchmark with 240 experimentally measured ("high-quality") and over 100K simulated (imperfect or "low-quality") HEA yield strength annotations. Due to the scarcity of experimental annotations and the quality gap in imperfectly simulated data, existing transfer learning methods cannot generalize well on our dataset. We address this cross-quality few-shot transfer problem by leveraging model sparsification "twice" --- as a noise-robust feature learning regularizer at the pre-training stage, and as a data-efficient learning regularizer at the few-shot transfer stage. While the workflow already performs decently with ad-hoc sparsity patterns tuned independently for either stage, we take a step further by proposing a bi-level optimization framework termed Bi-RPT, that jointly learns optimal masks and automatically allocates sparsity levels for both stages. The optimization problem is solved efficiently using gradient unrolling, which is seamlessly integrated with the training process. The effectiveness of Bi-RPT is validated through extensive experiments on our new challenging X-Yield dataset, alongside other synthesized testbeds. Specifically, we achieve an 8.9~19.8% reduction in terms of the test mean squared error and 0.98~1.53% in terms of test accuracy, merely using 5-10% of the experimental data. Codes and sample data are in the supplement.

Electronic structure calculations based on Kohn-Sham density functional theory (KSDFT) that incorporate exact-exchange or hybrid functionals are associated with a large computational expense, a consequence of the inherent cubic scaling bottleneck and large associated prefactor, which limits the length and time scales that can be accessed. Although orbital-free density functional theory (OFDFT) calculations scale linearly with system size and are associated with a significantly smaller prefactor, they are limited by the absence of accurate density-dependent kinetic energy functionals. Therefore, the development of accurate density-dependent kinetic energy functionals is important for OFDFT calculations of large realistic systems. To this end, we propose a method to train kinetic energy functional models at the exact-exchange level of theory by using a dictionary of physically relevant terms that have been proposed in the literature in conjunction with linear or nonlinear regression methods to obtain the fitting coefficients. For our dictionary, we use a gradient expansion of the kinetic energy nonlocal models proposed in the literature and their nonlinear combinations, such as a model that incorporates spatial correlations between higher order derivatives of electron density at two points. The predictive capabilities of these models are assessed by using a variety of model one-dimensional (1D) systems that exhibit diverse bonding characteristics, such as a chain of eight hydrogens, LiF, LiH, C4H2, C4N2, and C3O2. We show that by using the data from model 1D KSDFT calculations performed using the exact-exchange functional for only a few neutral structures, it is possible to generate models with high accuracy for charged systems and electron and kinetic energy densities during self-consistent field iterations. In addition, we show that it is possible to learn both the orbital dependent terms, i.e., the kinetic energy and the exact-exchange energy, and models that incorporate additional nonlinearities in spatial correlations, such as a quadratic model, are needed to capture subtle features of the kinetic energy density that are present in exact-exchange-based KSDFT calculations.

The exponential growth and success of machine learning (ML) has resulted in its application in all scientific domains including material science. Advancement in experimental techniques has led to an increase in the volume of material science data encouraging material scientists to investigate data-driven solutions to scientific problems. While the resources available to get started with ML are ever increasing, there is little literature on traversing through the space of decisions that need to be made to implement a robust and trustworthy ML solution. A lack of such resources leads to researchers wading through articles and papers trying to determine the best approach for their problem and sometimes also falling prey to pitfalls in a real-world scenario. This paper aims to act as a guide for researchers who want to strategically approach a ML solution to their problem through the use of domain knowledge and systematic evaluation of the major aspects of a ML pipeline. We focus on four aspects of the ML pipeline: (1) problem formulation, (2) data curation, (3) feature representation and model selection, and (4) model generalizability and real-world performance. In each case, we discuss the space of decisions, provide examples from scientific literature, and illustrate how different choices can affect the outcome through a case study of predicting compressive strength of uniaxially pressed molecular solid, 2,4,6-triamino-1,3,5-trinitrobenzene (TATB) samples. Using a similar approach of critical thinking along with rigorous evaluation and diagnostics, researchers can be assured of the reliability of predictions from their ML models.

Machine learning models are increasingly used in materials studies because of their exceptional accuracy. However, the most accurate machine learning models are usually difficult to explain. Remedies to this problem lie in explainable artificial intelligence (XAI), an emerging research field that addresses the explainability of complicated machine learning models like deep neural networks (DNNs). This article attempts to provide an entry point to XAI for materials scientists. Concepts are defined to clarify what explain means in the context of materials science. Example works are reviewed to show how XAI helps materials science research. Challenges and opportunities are also discussed.

To achieve a secure interconnected Industrial Control System (ICS) architecture, security practitioners depend on accurate identification of network host behavior. However, accurate machine learning based host identification methods depends on the availability of significant quantities of network traffic data, which can be difficult to obtain due to system constraints such as network security, data confidentiality, and physical location. In this work, we propose a network traffic feature prediction method based on a generative model, which achieves high host identification accuracy. Furthermore, we develop a joint training algorithm to improve host identification performance compared to separate training of the generative model and the classifier responsible for host identification.

The absence of a reliable formulation of the kinetic energy density functional has hindered the development of orbital free density functional theory. Using the data-aided learning paradigm, we propose a simple prescription to accurately model the kinetic energy density of any system. Our method relies on a dictionary of functional forms for local and nonlocal contributions, which have been proposed in the literature, and the appropriate coefficients are calculated via a linear regression framework. To model the nonlocal contributions, we explore two new nonlocal functionals-a functional that captures fluctuations in electronic density and a functional that incorporates gradient information. Since the analytical functional forms of the kernels present in these nonlocal terms are not known from theory, we propose a basis function expansion to model these seemingly difficult nonlocal quantities. This allows us to easily reconstruct kernels for any system using only a few structures. The proposed method is able to learn kinetic energy densities and total kinetic energies of molecular and periodic systems, such as H2, LiH, LiF, and a one-dimensional chain of eight hydrogens using data from Kohn-Sham density functional theory calculations for only a few structures.

Characterization of modern cyber–physical Industrial Control System (ICS) devices is critical to the evaluation of their security posture and an understanding of the underlying industrial processes with which they interact. In this work, we address two related ICS device identification tasks: (1) separating ICS from non-ICS devices and (2) identifying specific ICS device types. We propose two distinct methods (one based on the existing IP2Vec method, and a novel traffic-features-based method) for achieving the first task. For transferability of the first task between two datasets, the traffic-features-based method performs significantly better (75% overall accuracy) compared to IP2Vec (22.5% overall accuracy). We further propose a novel method called DNP2Vec to address the second task. DNP2Vec is evaluated on two different datasets and achieves perfect multi-class classification accuracy (100%) for both datasets.

We recently developed a deep learning method that can determine the critical peak stress of a material by looking at scanning electron microscope (SEM) images of the material's crystals. However, it has been somewhat unclear what kind of image features the network is keying off of when it makes its prediction. It is common in computer vision to employ an explainable AI saliency map to tell one what parts of an image are important to the network's decision. One can usually deduce the important features by looking at these salient locations. However, SEM images of crystals are more abstract to the human observer than natural image photographs. As a result, it is not easy to tell what features are important at the locations which are most salient. To solve this, we developed a method that helps us map features from important locations in SEM images to non-abstract textures that are easier to interpret.

Machine-learning (ML) techniques hold the potential of enabling efficient quantitative micrograph analysis, but the robustness of ML models with respect to real-world micrograph quality variations has not been carefully evaluated. We collected thousands of scanning electron microscopy (SEM) micrographs for molecular solid materials, in which image pixel intensities vary due to both the microstructure content and microscope instrument conditions. We then built ML models to predict the ultimate compressive strength (UCS) of consolidated molecular solids, by encoding micrographs with different image feature descriptors and training a random forest regressor, and by training an end-to-end deep-learning (DL) model. Results show that instrument-induced pixel intensity signals can affect ML model predictions in a consistently negative way. As a remedy, we explored intensity normalization techniques. It is seen that intensity normalization helps to improve micrograph data quality and ML model robustness, but microscope-induced intensity variations can be difficult to eliminate.

Nanomaterials of varying compositions and morphologies are of interest for many applications from catalysis to optics, but the synthesis of nanomaterials and their scale-up are most often time-consuming and Edisonian processes. Information gleaned from scientific literature can help inform and accelerate nanomaterials development, but again, searching the literature and digesting the information are time-consuming manual processes for researchers. To help address these challenges, we developed scientific article-processing tools that extract and structure information from the text and figures of nanomaterials articles, thereby enabling the creation of a personalized knowledgebase for nanomaterials synthesis that can be mined to help inform further nanomaterials development. Starting with a corpus of ca. 35k nanomaterials-related articles, we developed models to classify articles according to the nanomaterial composition and morphology, extract synthesis protocols from within the articles’ text, and extract, normalize, and categorize chemical terms within synthesis protocols. We demonstrate the efficiency of the proposed pipeline on an expert-labeled set of nanomaterials synthesis articles, achieving 100% accuracy on composition prediction, 95% prediction on morphology prediction, 0.99 AUC on protocol identification, and up to 0.87 F1-score on chemical entity recognition. In addition to processing articles’ text, microscopy images of nanomaterials within articles are also automatically identified and analyzed to determine nanomaterials’ morphologies and size distributions. To enable users to easily explore the database, we developed a complementary browser-based visualization tool that provides flexibility in comparing across subsets of articles of interest. We use these tools and information to identify trends in nanomaterials synthesis, such as the correlation of certain reagents with various nanomaterial morphologies, which is useful in guiding hypotheses and reducing the potential parameter space during experimental design.

In this paper, we address the challenge of analyzing simulation data on HPC systems by using Apache Spark, which is a Big Data framework. One of the main problems we encountered with using Spark on HPC systems is the ephemeral data explosion, which is brought about by the curse of persistence in the Spark framework. Data persistence is essential in reducing I/O, but it comes at the cost of storage space. We show that in some cases, Spark scratch data can consume an order of magnitude more space than the input data being analyzed, leading to fatal out-of-disk errors. We investigate the real-world application of scaling machine learning algorithms to predict and analyze failures in multi-physics simulations on 76TB of data (over one trillion training examples). This problem is 2--3 orders of magnitude larger than prior work. Based on extensive experiments at scale, we provide several concrete recommendations as state-of-the-practice, and demonstrate a 7x reduction in disk utilization with negligible increases or even decreases in runtime.

We explore the application of computer vision and machine learning (ML) techniques to predict material properties (e.g. compressive strength) based on SEM images. We show that it's possible to train ML models to predict materials performance based on SEM images alone, demonstrating this capability on the real-world problem of predicting uniaxially compressed peak stress of consolidated molecular solids samples. Our image-based ML approach reduces mean absolute percent error (MAPE) by an average of 24% over baselines representative of the current state-of-the-practice (i.e., domain-expert's analysis and correlation). We compared two complementary approaches to this problem: (1) a traditional ML approach, random forest (RF), using state-of-the-art computer vision features and (2) an end-to-end deep learning (DL) approach, where features are learned automatically from raw images. We demonstrate the complementarity of these approaches, showing that RF performs best in the "small data" regime in which many real-world scientific applications reside (up to 24% lower RMSE than DL), whereas DL outpaces RF in the "big data" regime, where abundant training samples are available (up to 24% lower RMSE than RF). Finally, we demonstrate that models trained using machine learning techniques are capable of discovering and utilizing informative crystal attributes previously underutilized by domain experts.

The Arbitrary Lagrangian-Eulerian (ALE) method is used in a variety of engineering and scientific applications for enabling multi-physics simulations. Unfortunately, the ALE method can suffer from failures that require users to adjust a set of parameters to control mesh relaxation. In this paper, we present a deep learning framework for predicting mesh relaxation in ALE simulations. Our framework is designed to train a neural network using data generated from existing ALE simulations developed by expert users. In order to capture the spatial coherence inherent in simulations, we apply convolutional-deconvolutional neural networks to achieve up to 0.99 F1 score in predicting mesh relaxation.

Despite ML’s impressive performance in commercial applications, several unique challenges exist when applying ML in materials science applications. In such a context, the contributions of this work are twofold. First, we identify common pitfalls of existing ML techniques when learning from underrepresented/imbalanced material data. Specifically, we show that with imbalanced data, standard methods for assessing quality of ML models break down and lead to misleading conclusions. Furthermore, we find that the model’s own confidence score cannot be trusted and model introspection methods (using simpler models) do not help as they result in loss of predictive performance (reliability-explainability trade-off). Second, to overcome these challenges, we propose a general-purpose explainable and reliable machine-learning framework. Specifically, we propose a generic pipeline that employs an ensemble of simpler models to reliably predict material properties. We also propose a transfer learning technique and show that the performance loss due to models’ simplicity can be overcome by exploiting correlations among different material properties. A new evaluation metric and a trust score to better quantify the confidence in the predictions are also proposed. To improve the interpretability, we add a rationale generator component to our framework which provides both model-level and decision-level explanations. Finally, we demonstrate the versatility of our technique on two applications: (1) predicting properties of crystalline compounds and (2) identifying potentially stable solar cell materials. We also point to some outstanding issues yet to be resolved for a successful application of ML in material science.

Networks represent relationships between entities in many complex systems, spanning from online social interactions to biological cell development and brain connectivity. In many cases, relationships between entities are unambiguously known: are two users "friends" in a social network? Do two researchers collaborate on a published article? Do two road segments in a transportation system intersect? These are directly observable in the system in question. In most cases, relationships between nodes are not directly observable and must be inferred: Does one gene regulate the expression of another? Do two animals who physically co-locate have a social bond? Who infected whom in a disease outbreak in a population? Existing approaches for inferring networks from data are found across many application domains and use specialized knowledge to infer and measure the quality of inferred network for a specific task or hypothesis. However, current research lacks a rigorous methodology that employs standard statistical validation on inferred models. In this survey, we examine (1) how network representations are constructed from underlying data, (2) the variety of questions and tasks on these representations over several domains, and (3) validation strategies for measuring the inferred network's capability of answering questions on the system of interest.