High-energy diffraction microscopy (HEDM) is a valuable technique that allows for time-resolved 3D characterization of grain-scale deformation during mechanical testing. In past studies, the local spatially resolved information provided by HEDM has revealed unintentionally imposed bending moments that complicate data analysis and efforts to couple experiments with microstructure-sensitive models. In this work, a mechanical test and in-situ HEDM experiment with both bending and tension loads applied was conducted on a polycrystalline nickel-based superalloy, IN718. Crystal plasticity finite element (CPFE) simulations were performed using the experimental microstructure as input and loaded to reproduce the bending-tension loads, as well as a pure tension loading procedure. This study examines the evolution of grain-by-grain strain and intragranular orientation distributions in both the experiment and model throughout loading, thus revealing the persistent effect of bending in both macroscopic elasticity and plasticity. Additionally, CPFE predictions of grains that accumulate relatively high and low amounts of slip reveal that extreme plasticity value grains can be sensitive to bending. Finally, recommendations are given for the interpretation and modeling of bending moment containing HEDM experiments.
Material properties are strongly influenced by their microstructure. Thus, the microstructure must be accurately quantified to predict material behavior. Although these properties are governed by their three-dimensional (3D) microstructure, most common characterization techniques rely on two-dimensional (2D) imaging. Serial sectioning methods can accurately capture 3D microstructural features, but they are time-consuming, expensive, and destructive. This work addresses 3D microstructure characterization using a power-law mapping derived from 2D images. To this end, a dataset of synthetic microstructures with equiaxed grains and six different grain size distributions is used. The extracted grain size distributions highlight the differences between 2D and 3D statistics. The study also quantifies the effect of the number of grains within a 2D image on the resulting statistics. The power-law mapping is also compared against the Saltykov stereological method. The mapping outperforms the Saltykov method for number-weighted statistics, where the Saltykov method exhibits a systematic underestimation of the mean of approximately 27
The quality of additively manufactured (AM) polymers-based nanocomposites is governed by material composition and processing conditions, and can be predicted by statistical machine learning (ML) models. However, incorporating material features requires extensive material characterization. This study proposes a systematic data-driven workflow for printability prediction in extrusion-based additive manufacturing built around a deliberately simplified characterization strategy centered on capillary rheometry. Representative polymers and their nanocomposites exhibiting diverse printability were characterized via rotational and capillary rheometry as well as thermal analysis. Extruded pellets were later printed at varying temperatures and flow rates, and key printability and dimensional metrics were evaluated against three target variables: variation in mass ( Δ W ), variation in internal diameter ( Δ D_i ) and surface roughness (RA). Input feature sets spanning capillary-derived parameters, oscillatory rheological descriptors, thermal properties, and process conditions were systematically constructed and compared across classification and regression tasks. A minimal proxy set composed solely of capillary-derived Carreau parameters and process conditions achieved 81
Machine learning is explored to assess optical micrographs for defects and in identifying machine settings of laser-based powder bed fusion additively manufactured (AM) steel. Two algorithms are explored for defect analysis: Naïve Bayes and random forest. Additionally, a novel technique for discriminating density analysis of optical micrographs of highly dense builds is developed. These tools aim to fill gaps in both rapid qualification of AM components and allow for faster, more accurate parameter sweeps for development of optimal printing parameters for new materials. The Naïve Bayes algorithm outperforms the random forest algorithm in this study because it allows for less rigidly linear classification subsets. The novel gray thresholding routine is compared to all available automatic thresholding methods available in freeware ImageJ and is shown to discriminate between high density optical micrographs ( ≥ 99.5
Artificial intelligence is entering materials science and manufacturing at a moment when retrieval is often mistaken for discovery. We argue that models trained to optimize plausibility within a closed theoretical system are structurally biased toward interpolation and away from the anomalies from which scientific revolutions arise. In materials and manufacturing, this limitation becomes acute because the design space is combinatorial, multiscale, and physically constrained. What is needed is not another assistant for summarizing what is already known, but systems that can assemble compositional world models, invert the forward problem, falsify their own principles, and test those principles against simulation and fabrication. Drawing on recent work in graph-native reasoning, inverse protein design, multi-agent scientific discovery, and AI-integrated manufacturing, we outline an architecture for discovery built on three coupled capabilities: world-model construction, adversarial falsification, and physical grounding. We make this architecture operational through a minimum description length gate: A proposed world-model break is accepted only when the revised model encodes the accumulated evidence in fewer bits than the model it replaces. We illustrate the loop with a residue-level protein mechanics case study in which a Breaker–Builder agent system revises a symbolic graph model of crystallographic B-factor data through accepted, rejected, and retracted hypotheses. The central claim is that materials innovation will accelerate when AI moves beyond retrieval and surrogate prediction toward systems that can generate, challenge, and physically realize new hypotheses. In that sense, discovery requires breaking the current world model in order to build the next one.
X-ray Micro Computed Tomography (X-µCT) is increasingly regarded as the gold standard for inspecting additively manufactured components used in fatigue-critical applications. However, segmentation of X-µCT data remains inconsistent across users and applications. Additionally, it is unclear if voxel-wise metrics of segmentation quality, such as the Dice coefficient or Intersection over Union (IoU), are relevant to fatigue performance. In this work, we evaluated global binary thresholding, adaptive thresholding, hysteresis thresholding, and a 2.5D U-Net on X-µCT scans of Powder Bed Fusion – Laser Beam manufactured Ti-6Al-4V rotating bending fatigue specimens from the NIST AMBench 2025 challenge (AMB2025-03-FL) to quantify segmentation-induced measurement bias and assess its impact on predicting the fatigue-initiating pore. To identify the fatigue-initiating pore, the Murakami √(A) parameter was modified using beam theory to account for the stress gradient due to bending. This modification enables localization of the most critical pore, independent of the applied stress. Using the modified √(A) metric, the global binary thresholding, adaptive thresholding, and hysteresis thresholding predicted the fatigue-initiating pore correctly for three out of the four test cases. The 2.5D U-Net was able to predict the fatigue-initiating pore in all four cases despite having a lower Dice coefficient and IoU values when compared to the other segmentation models. In fatigue-critical applications, these limitations are most consequential for small, near-surface pores, where X-ray reflection artifacts cause threshold-based methods to underestimate pore size. Such pores can occur even in high-density PBF-LB parts. Consequently, voxel-wise metrics such as Dice coefficient or IoU do not indicate whether a segmentation approach can identify the true fatigue-initiating pore, as they weight all pixels equally, highlighting the need for fatigue-aware, feature-based methods for segmentation of X-µCT data.
Identifying the depth of an acoustic emission (AE) source in plate-like geometries is an open problem with significant practical implications for understanding ply-level damage evolution in composites. Currently, these characterizations are limited to half the plate thickness, as sources symmetric about the plate midline produce similar signals. The challenge arises primarily because AE signals are analyzed using traditional low-dimensional features, such as peak amplitude or peak frequency, which exhibit little distinction across source depths and thereby prevent unique depth identification. In this work, we demonstrate that statistical learning models, trained on high-dimensional feature vectors, can overcome this limitation using two benchmark datasets (1680 waveforms each) of pencil-lead breaks (PLBs) performed at the top, side, and bottom of an aluminum plate. To further simulate realistic conditions, the datasets were designed to isolate sensor coupling and source-to-sensor distance effects, respectively. Models trained on low-dimensional features (<12 parameters) perform poorly at distinguishing top from bottom PLBs. Notably, models derived from the distribution of energy in the frequency spectra achieve 90
This work presents a novel ABAQUS-Voronoi toolbox for generating polycrystalline representative volume elements (RVEs) specifically designed for the modeling of thin metal sheets. Developed in Python and fully integrated into the ABAQUS/CAE Graphical User Interface (GUI), the toolbox introduces a key innovation, which consists in the direct application of Voronoi-like partition to pre-existing finite element meshes. The proposed toolbox provides two main modules: a general module for rapid RVE generation and a user-defined module for importing external seeds. This toolbox enables the construction of RVEs with complex features, such as voids, non-cubic geometries, and gradient or columnar grain morphologies. The toolbox performance is demonstrated within a multiscale framework based on the Crystal Plasticity Finite Element Method (CPFEM). Results confirm its flexibility and robustness, showing that the generated RVEs can successfully predict the overall trends of the macroscopic mechanical response of polycrystalline aggregates.
This work characterizes the processing–structure–property relationship of field-processed particle polymer matrix composites. Electric and magnetic fields coerce particles to form structures that drive effective properties, which are locked in once the matrix material is cured or solidified. In combination with additive manufacturing, field-processed composites allow for tuning of local properties throughout a printed component. Effectively designing with these materials requires an understanding of the effects of the processing on the bulk material properties. This work introduces a framework to determine the resulting structures formed with applied fields, demonstrated with 5 base cases (no applied field, electric field only, magnetic field only, electric and magnetic field perpendicular, and electric and magnetic field parallel). The conditions are modeled using a representative volume element (RVE) and particle dynamics simulations in MATLAB. Resulting structures are then homogenized using COMSOL Multiphysics to predict effective properties. Simulations are conducted for 1, 5, and 10
Externally solidified crystals (ESCs) can easily form during high pressure die casting (HPDC) of aluminum alloys if the process is not properly controlled, leading to significant reduction of casting quality and performance. This study explores seven machine learning (ML) models, including decision tree, random forest, logistic regression, neural network, K-nearest neighbors (KNN), support vector machine (SVM), and naïve Bayes classifier. The random forest and classification tree models showed the highest accuracy of 95
The materials science literature is the richest reservoir of domain knowledge, yet converting its unstructured text—especially narrative passages and complex tables—into machine-readable data for analysis and machine learning (ML) model training remains challenging. To address this, we present KnowMat, an agentic, multistage pipeline that transforms full-text articles into schema-aligned, machine-readable JSON. KnowMat parses PDFs (text and tables) and performs iterative extraction with evaluation-driven re-runs to enhance coverage while curbing hallucinations. A two-stage manager then aggregates, validates, and corrects results, while properties are encoded with a fidelity-preserving dual representation (original textual form along with numeric surrogate with explicit value type); standardized labels are added without altering author-reported names to support database integration. Although demonstrated for materials literature, the workflow is schema-agnostic and readily adaptable to other scientific domains. Evaluation on real-world materials science papers demonstrates KnowMat’s accuracy and efficiency, significantly reducing barriers to data-driven materials research.
Experimental evaluation demonstrates strong performance, achieving an F1 score of 0.92 for worker detection and 0.93 for personal protective equipment (PPE) classification. The integrated hazard-zone monitoring module further enables spatial safety enforcement, achieving a frame-level detection accuracy of 92
Manual quantitative analysis of large microstructural datasets is a challenging process. To address this, we present a novel pipeline for automated microstructure segmentation and 3D reconstruction in aluminum alloys, combining Mask R-CNN with a custom reconstruction algorithm. Datasets generated from light optical microscopy (LOM), computed tomography (CT) and phase field simulations (PFS) were used to train and evaluate the Mask R-CNN deep learning model. The alloys investigated include AlSi6Cu4Fe1, AlSi6Cu4Fe2 and AlCu10. The microstructures in sections comprise needle-like intermetallic precipitates and irregularly shaped precipitates in the former two alloys, and columnar dendrites formed during directional solidification in the latter. A 3D reconstruction algorithm was developed to generate three-dimensional representations from CT slices of individual dendrites and full monolithic structures of interconnected precipitates based on the Mask R-CNN detections. This reconstruction algorithm was validated using a 3D dataset from phase-field simulations to ensure accuracy and reliability. The deep learning model consistently achieved high detection accuracy across LOM and CT datasets for all investigated microstructural objects, reaching average accuracy of 73
Establishing reliable correlations among material properties across different scales is essential for enabling informed materials selection, performance estimation, and property screening. While most existing datasets and modeling efforts focus on predicting individual properties from composition or microstructure, limited attention has been given to uncovering the interrelations between modulus-related, thermal, and strength properties. Although empirical relationships have been proposed to relate certain properties, these are limited to specific alloys and fail to generalize across scales. To address the challenge, this paper leverages machine learning (ML) to uncover hidden nonlinear relationships between properties. Firstly, 1731 experimentally validated alloys were curated from the ANSYS GRANTA database, and materials properties were categorized into thermally, mechanically, and strength-related groups. Three ML models are employed to learn relationships among alloy properties: neural networks (NN), geometric harmonics (GH), and double diffusion maps (DDM). Results demonstrate that thermal and mechanical properties are strongly interrelated and can be predicted with high accuracy. However, the strength property is revealed to be difficult to model due to the missing information on post-processing treatments. To address this, we incorporated treatment metadata for iron-based alloys. The inclusion of these data led to an increase of over 70
Only a limited number of alloys can be additively manufactured industrially today, to form defect-free parts. Optimization of process parameters of laser power, scan speed, powder layer thickness, and hatch spacing during laser-based additive manufacturing, particularly laser powder bed fusion (LPBF), is of prime importance to take advantage of unique microstructures and enhanced mechanical properties of these 3D-printed parts. This study presents a physics-informed, data-driven framework for predicting defect formation and constructing printability maps in LPBF of industrial alloys. By integrating solidification-coupled computational fluid dynamics (CFD) simulations, single-track experiments, and machine learning classification, the approach captures the effect of melt-pool flow characteristics and material thermophysical properties on defect formation. The resulting printability maps accurately delineate stable, balling, lack of fusion, and keyhole regimes across SS316L, IN718, and AlSi10Mg alloys. SHapley Additive exPlanation (SHAP)-based model interpretation reveals that process energy input and alloy properties jointly govern defect transitions, aligning with established LPBF physics. The framework offers a generalizable route for identifying defect-free processing windows in new alloys, reducing empirical optimization efforts and accelerating qualification for additive manufacturing applications.
Additive manufacturing enables the production of metal parts with complex geometries and advanced capabilities. However, the progress of additive manufacturing in producing multifunctional materials with solid-state energy conversion capability is still limited. To use additive manufacturing on multifunctional materials, it is essential to understand and predict the microstructure resulting from additive manufacturing’s unique processing conditions. This work studies bismuth telluride (Bi2Te3), a well-known semiconductor thermoelectric material, undergoing a laser powder bed fusion process. The process-structure relationship is investigated both experimentally and computationally. The grain structure formation over a wide range of processing parameters is captured computationally, with finite element modeling and kinetic Monte Carlo simulations, and experimentally, with single melt line studies on Bi2Te3. This work compares the process-structure relationship of bismuth telluride to that of well-studied metals to understand how additive manufacturing may differ for alloys like thermoelectric semiconductor materials. The melt pool processed experimentally under 25 W and 400 mm/s processing parameters was fully melted and exhibited conduction melting mode. The computational methodology demonstrated strong predictive capability for grain size and distribution in conduction mode melting but was less accurate in modes featuring Marangoni convection and keyhole formation. Laser power and laser scan speed resulting in conduction melting mode for bismuth telluride are lower than for metals like stainless steel, Inconel, and magnesium; the lower values compensate for differences in thermal conductivity, resulting in comparable temperature gradients. The solidification rate for bismuth telluride is lower than for metals due to slower laser scan speeds. Below is the link to the editor’s video summary.
This work presents an original surrogate model for the accurate and computationally efficient prediction of molten pool size in multi-track laser melting over a large domain at operating conditions relevant to laser powder bed fusion. While high-fidelity models can accurately predict the molten pool dynamics, the high computational expense limits their application to a few short tracks on small domains. Conduction models, on the other hand, are orders of magnitude cheaper to evaluate but lack the necessary physics for accurate predictions. This research presents a surrogate model that combines the computational efficiency of the conduction model with the accuracy of the high-fidelity model. A conduction model and high-fidelity model are simulated over a small scan pattern to generate training data of the highly transient molten pool depth and width. A data-driven model, consisting of a fuzzy basis function network, is trained with the aforementioned data. The conduction model is then simulated over a larger scan pattern, the results are input into the trained surrogate model, thereby outputting high-fidelity predictions of the molten pool size over a larger scan pattern. Comparison with experimental results shows this surrogate modeling framework provides reasonably accurate predictions of the molten pool depth and width (average error of -5.2
Incremental hot forging is a manufacturing process with a long history and continued prevalence today. This process involves repeated heating and compressing of a metal workpiece to deform the material to the desired geometry. The processing of metals in this way yields a final part crafted from a single piece of material that leads to desirable properties such as homogeneous microstructure and low porosity. The gold standard for predicting end-state geometry is finite element analysis (FEA). FEA is accurate as the physics of deformation are approximated, but available FEA tools require significant processing time. In this work, we detail the use of a graph neural network (GNN) to provide a method for rapid workpiece shape prediction based solely on surface geometry node positions and edge connectivity of the workpiece and press. The motivation for this work is to demonstrate a proof-of-concept method that can generate predicted shape information in a fraction of the time required for FEA. Our approach demonstrates learning toward the prediction of forged shape based on a limited set of training data from prior FEA simulations and proves to be effective for decreasing the time required for shape prediction.
Ever-increasing environmental damage and health concerns associated with conventional petroleum-based plastics pollution have sparked a global drive toward sustainable alternatives in packaging domain. Biopolymers, sourced from renewable biological materials such as microorganisms, plant and agro waste, and marine and animal by-products have emerged as promising candidates for developing sustainable packaging solutions due to their ability to be tweaked for desirable functional properties along with biodegradability. However, major challenges persist in aligning material change with explicit consumer requirements with prioritization to desired properties based on functional and non-functional expectations, cost, and consistent supply without compromising environmental goals. Consequently, there is a pressing need for integrated approaches that reconcile material performance, economic feasibility, and environmental sustainability for practical change. In this paper, we propose the PFS (performance–feasibility–sustainability) model for early-stage material selection for packaging that provides flexibility in parameter arrangements for better customer requirements, optimizing performance, supply chain feasibility, and safety assessments of the biopolymers. We elucidate with an example of replacing automobile car seat fabric normally covered with polyethylene with a flexible sustainable biopolymer. Customer requirements, their preferences for each category, along with the underlying parameters provided by domain specialists, are accurately aligned with our biopolymer database. We execute the PFS model to obtain the desired result using various weighed sum approaches like Weighted Sum Model (WSM), Technique for Order Preference by Similarity to Ideal Solution (TOPSIS), and Quotient model. WSM suggested lignin as first choice of biopolymer for bio packaging. Polycaprolactone was recommended by the TOPSIS and Quotient model as output with the highest PFS Score. This is our attempt to develop a quantifiable agile methodology for early-stage material selection from a range of biopolymers intended for packaging applications.
Hot isostatic pressing (HIP) is a critical post-processing step to ensure the reliability and performance of additively manufactured (AM) niobium (Nb) components for demanding aerospace and high-temperature applications. However, optimizing HIP cycles to maximize strength without relying on extensive and costly physical trials presents a significant industrial challenge. This work adapts a predictive phase-field modeling framework that directly correlates HIP process parameters of temperature and pressure to the final microstructure and mechanical strength of AM Nb, serving as a powerful tool for virtual process design. The simulation results provide quantitative guidance for manufacturing process design. The model demonstrates that applied pressure is a key lever for suppressing grain growth at high temperatures, thereby enhancing component strength. For instance, at a processing temperature of 1373 K, increasing the HIP pressure from 1 MPa to 100 MPa is predicted to boost the final yield strength by 5.70
