Industrial materials images are an important application domain for content-based image retrieval. Users need to quickly search databases for images that exhibit similar appearance, properties, and/or features to reduce analysis turnaround time and cost. The images in this study are 2D images of millimeter-scale rock samples acquired at micrometer resolution with light microscopy or extracted from 3D micro-CT scans. Labeled rock images are expensive and time-consuming to acquire and thus are typically only available in the tens of thousands. Training a high-capacity deep learning (DL) model from scratch is therefore not practicable due to data paucity. To overcome this “few-shot learning” challenge, we propose leveraging pretrained common DL models in conjunction with transfer learning. The “similarity” of industrial materials images is subjective and assessed by human experts based on both visual appearance and physical qualities. We have emulated this human-driven assessment process via a physics-informed neural network including metadata and physical measurements in the loss function. We present a novel DL architecture that combines Siamese neural networks with a loss function that integrates classification and regression terms. The networks are trained with both image and metadata similarity (classification), and with metadata prediction (regression). For efficient inference, we use a highly compressed image feature representation, computed offline once, to search the database for images similar to a query image. Numerical experiments demonstrate superior retrieval performance of our new architecture compared with other DL and custom-feature-based approaches.

The decreasing cost and improved sensor and monitoring system technology (e.g., fiber optics and strain gauges) have led to more measurements in close proximity to each other. When using such spatially dense measurement data in Bayesian system identification strategies, the correlation in the model prediction error can become significant. The widely adopted assumption of uncorrelated Gaussian error may lead to inaccurate parameter estimation and overconfident predictions, which may lead to suboptimal decisions. This article addresses the challenges of performing Bayesian system identification for structures when large datasets are used, considering both spatial and temporal dependencies in the model uncertainty. We present an approach to efficiently evaluate the log-likelihood function, and we utilize nested sampling to compute the evidence for Bayesian model selection. The approach is first demonstrated on a synthetic case and then applied to a (measured) real-world steel bridge. The results show that the assumption of dependence in the model prediction uncertainties is decisively supported by the data. The proposed developments enable the use of large datasets and accounting for the dependency when performing Bayesian system identification, even when a relatively large number of uncertain parameters is inferred.

The finite element method (FEM) is widely used to simulate a variety of physics phenomena. Approaches that integrate FEM with neural networks (NNs) are typically leveraged as an alternative to conducting expensive FEM simulations in order to reduce the computational cost without significantly sacrificing accuracy. However, these methods can produce biased predictions that deviate from those obtained with FEM, since these hybrid FEM-NN approaches rely on approximations trained using physically relevant quantities. In this work, an uncertainty estimation framework is introduced that leverages ensembles of Bayesian neural networks to produce diverse sets of predictions using a hybrid FEM-NN approach that approximates internal forces on a deforming solid body. The uncertainty estimator developed herein reliably infers upper bounds of bias/variance in the predictions for a wide range of interpolation and extrapolation cases using a three-element FEM-NN model of a bar undergoing plastic deformation. This proposed framework offers a powerful tool for assessing the reliability of physics-based surrogate models by establishing uncertainty estimates for predictions spanning a wide range of possible load cases.

In this study, we present and assess data-driven approaches for modeling contact line dynamics, using droplet transport on chemically heterogeneous surfaces as a model system. Ground-truth data for training and validation are generated based on long-wave models that are applicable for slow droplet motion with small contact angles, which are known to accurately reproduce the dynamics with minimal computing resources compared to high-fidelity direct numerical simulations. The data-driven models are based on the Fourier neural operator (FNO) and are developed following two different approaches. The first deploys the data-driven method as an iterative neural network architecture, which predicts the future state of the contact line based on a number of previous states. The second approach corrects the time derivative of the contact line by augmenting its low-order asymptotic approximation with a data-driven counterpart, evolving the resulting system using standard time integration methods. The performance of each approach is evaluated in terms of accuracy and generalizability, concluding that the latter approach, although not originally explored within the original contribution on the FNO, outperforms the former.

In the present work, neural networks are applied to formulate parametrized hyperelastic constitutive models. The models fulfill all common mechanical conditions of hyperelasticity by construction. In particular, partially input convex neural network (pICNN) architectures are applied based on feed-forward neural networks. Receiving two different sets of input arguments, pICNNs are convex in one of them, while for the other, they represent arbitrary relationships which are not necessarily convex. In this way, the model can fulfill convexity conditions stemming from mechanical considerations without being too restrictive on the functional relationship in additional parameters, which may not necessarily be convex. Two different models are introduced, where one can represent arbitrary functional relationships in the additional parameters, while the other is monotonic in the additional parameters. As a first proof of concept, the model is calibrated to data generated with two differently parametrized analytical potentials, whereby three different pICNN architectures are investigated. In all cases, the proposed model shows excellent performance.

There have been consistent calls for more research on managing teams and embedding processes in data science innovations. Widely used frameworks (e.g., the cross-industry standard process for data mining) provide a standardized approach to data science but are limited in features such as role clarity, skills, and cross-team collaboration that are essential for developing organizational capabilities in data science. In this study, we introduce a data workflow method (DWM) as a new approach to break organizational silos and create a multi-disciplinary team to develop, implement and embed data science. Different from current data science process workflows, the DWM is managed at the system level that shapes business operating model for continuous improvement, rather than as a function of a particular project, one single business unit, or isolated individuals. To further operationalize the DWM approach, we investigated an embedded data workflow at a mining operation that has been using geological data in a machine-learning model to stabilize daily mill production for the last 2 years. Based on the findings in this study, we propose that the DWM approach derives its capability from three aspects: (a) a systemic data workflow; (b) multi-disciplinary networks of collaboration and responsibility; and (c) clearly identified data roles and the associated skills and expertise. This study suggests a whole-of-organization approach and pathway to develop data science capability.

Modeling complex dynamical systems with only partial knowledge of their physical mechanisms is a crucial problem across all scientific and engineering disciplines. Purely data-driven approaches, which only make use of an artificial neural network and data, often fail to accurately simulate the evolution of the system dynamics over a sufficiently long time and in a physically consistent manner. Therefore, we propose a hybrid approach that uses a neural network model in combination with an incomplete partial differential equations (PDEs) solver that provides known, but incomplete physical information. In this study, we demonstrate that the results obtained from the incomplete PDEs can be efficiently corrected at every time step by the proposed hybrid neural network—PDE solver model, so that the effect of the unknown physics present in the system is correctly accounted for. For validation purposes, the obtained simulations of the hybrid model are successfully compared against results coming from the complete set of PDEs describing the full physics of the considered system. We demonstrate the validity of the proposed approach on a reactive flow, an archetypal multi-physics system that combines fluid mechanics and chemistry, the latter being the physics considered unknown. Experiments are made on planar and Bunsen-type flames at various operating conditions. The hybrid neural network—PDE approach correctly models the flame evolution of the cases under study for significantly long time windows, yields improved generalization and allows for larger simulation time steps.

The curse of dimensionality confounds the comprehensive evaluation of computational structural mechanics problems. Adequately capturing complex material behavior and interacting physics phenomenon in models can lead to long run times and memory requirements resulting in the need for substantial computational resources to analyze one scenario for a single set of input parameters. The computational requirements are then compounded when considering the number and range of input parameters spanning material properties, loading, boundary conditions, and model geometry that must be evaluated to characterize behavior, identify dominant parameters, perform uncertainty quantification, and optimize performance. To reduce model dimensionality, global sensitivity analysis (GSA) enables the identification of dominant input parameters for a specific structural performance output. However, many distinct types of GSA methods are available, presenting a challenge when selecting the optimal approach for a specific problem. While substantial documentation is available in the literature providing details on the methodology and derivation of GSA methods, application-based case studies focus on fields such as finance, chemistry, and environmental science. To inform the selection and implementation of a GSA method for structural mechanics problems for a nonexpert user, this article investigates five of the most widespread GSA methods with commonly used structural mechanics methods and models of varying dimensionality and complexity. It is concluded that all methods can identify the most dominant parameters, although with significantly different computational costs and quantitative capabilities. Therefore, method selection is dependent on computational resources, information required from the GSA, and available data.

The increase in global population and urbanization is presenting significant challenges to society: space is becoming increasingly scarce, demand is exceeding capacity for deteriorating infrastructure, transportation is fraught with congestion, and environmental impacts are accelerating. Underground space, and particularly tunnels, has a key role to play in tackling these challenges. However, the cost, risk, uncertainty, and complexity of the tunneling process have impeded its growth. In this paper, we envision several technological advancements that can potentially innovate and transform the mechanized tunneling industry, including artificial intelligence (AI), autonomous, and bio-inspired systems. The proliferation of AI may assist human engineers and operators in making informed decisions systematically and quantitatively based on massive real-time data during tunneling. Autonomous tunneling systems may enable precise and predictable tunneling operations with minimal human intervention and facilitate the construction of massive and large-scale underground infrastructure projects that were previously challenging or unfeasible using conventional methods. Bio-inspired systems may provide valuable references and strategies for more efficient tunneling design and construction concepts. While these technological advancements can offer great promise, they also face considerable challenges, such as improving accessibility to and shareability of tunneling data, developing robust, reliable, and explainable machine learning systems, as well as scaling the mechanics and ensuring the applicability of bio-inspired systems from the prototype level to real-world applications. Addressing these challenges is imperative to ensure the successful implementation of these innovations for future tunneling.

Detection of defects and identification of symptoms in monitoring industrial systems is a widely studied problem with applications in a wide range of domains. Most of the monitored information extracted from systems corresponds to data series (or time series), where the evolution of values through one or multiple dimensions directly illustrates its health state. Thus, an automatic anomaly detection method in data series becomes crucial. In this article, we propose a novel method based on a convolutional neural network to detect precursors of anomalies in multivariate data series. Our contribution is twofold: We first describe a new convolutional architecture dedicated to multivariate data series classification; We then propose a novel method that returns dCAM, a dimension-wise Class Activation Map specifically designed for multivariate time series that can be used to identify precursors when used for classifying normal and abnormal data series. Experiments with several synthetic datasets demonstrate that dCAM is more accurate than previous classification approaches and a viable solution for discriminant feature discovery and classification explanation in multivariate time series. We then experimentally evaluate our approach on a real and challenging use case dedicated to identifying vibration precursors on pumps in nuclear power plants.