PaperlayerView exact revision on arXiv

Source-linked AI summary

Machine Learning for Materials Developments in Metals Additive Manufacturing

N. S. Johnson, P. S. Vulimiri, A. C. To, X. Zhang, C. A. Brice, B. B. Kappes, A. P. Stebner

arXiv:2005.05235v1physics.app-phcond-mat.mtrl-sciphysics.comp-ph

TL;DR

Metals AM must control many coupled material and design variables while maintaining consistent part quality, making development computationally and experimentally demanding. This review explains ML concepts and surveys their use alongside physics-based methods across AM design and process problems. It concludes that ML can accelerate design exploration and modeling, but only when applied to suitable data and problems.

  • Problem

    Metals AM simultaneously fabricates materials and components across tens to hundreds of coupled degrees of freedom, while consistency and quality remain open challenges.

  • Method

    The review introduces ML terminology and algorithms, then examines how ML can complement physics-based ICME across metals AM design and process applications.

  • Results

    Machine learning methods can produce topology predictions comparable in accuracy to traditional methods while greatly reducing computational time for design exploration.

  • Takeaways & Limitations

    ML is a complementary tool for integrated ICME modeling, data-driven design, and feedback and control using in situ process-monitoring sensors.

Abstract

from arXiv · show

In metals additive manufacturing (AM), materials and components are concurrently made in a single process as layers of metal are fabricated on top of each other in the near-final topology required for the end-use product. Consequently, tens to hundreds of materials and part design degrees of freedom must be simultaneously controlled and understood; hence, metals AM is a highly interdisciplinary technology that requires synchronized consideration of physics, chemistry, materials science, physical metallurgy, computer science, electrical engineering, and mechanical engineering. The use of modern machine learning approaches to model these degrees of freedom can reduce the time and cost to elucidate the science of metals AM and to optimize the engineering of these complex, multidisciplinary processes. New machine learning techniques are not needed for most metals AM development; those used in other sects of materials science will also work for AM. Most prolifically, the density functional theory (DFT) community has used many of them since the early 2000s for evaluating numerous combinations of elements and crystal structures to discover new materials. This materials technologies-focused review introduces the basic mathematics and terminology of machine learning through the lens of metals AM, and then examines potential uses of machine learning to advance metals AM, highlighting the many parallels to previous efforts in materials science and manufacturing while also discussing new challenges and adaptations specific to metals AM.

I. MOTIVATION

Metals AM fabricates materials and parts simultaneously, creating extensive process–structure–property–performance linkages and persistent consistency challenges. The review positions machine learning as a complement to physics-based ICME for addressing these complexities.

  • Metals AM simultaneously fabricates material and component geometry, enabling complex shapes and locally varied microstructure–property relationships through process changes.
  • Consistency and quality remain open challenges because AM parts can exhibit high microstructure variability and defect densities that contribute to unexpected fatigue failures.
  • AM process physics are difficult to model because rapid solidification, vaporization, volatile-element ingestion, and location-dependent thermal histories vary within and across builds.
  • Machine learning is presented as a way to couple physics-based ICME tools with experimental data and accelerate materials and process development.
  • The review introduces ML terminology and algorithms, surveys materials-science applications, and identifies uses and implementation obstacles specific to metals AM.

B. Data Sources

AM data span scalar measurements, distributions, series, categories, images, and specialized computational or geometric representations. Effective ML use requires converting these sources into informative features, curating their statistical structure, and organizing inputs and outputs appropriately.

  • AM datasets include machine parameters, material properties, distributions, time series, categorical descriptors, microscopy and process images, vector fields, texture functions, and 3D build representations.
  • Featurization extracts information that exposes relationships among features or between features and outcomes such as strength, roughness, or shape.
  • Statistical curation examines correlations, outliers, distributions, and redundant inputs because improper preprocessing or correlated features can degrade prediction or increase computation.
  • Matrices represent observations as rows and features as columns, while labeled datasets store process inputs separately from resulting property outputs for prediction.

D. The Assumptions Behind Machine Learning

Machine learning relies on relational and similarity assumptions about data and responses. The review distinguishes unsupervised learning, which finds trends, from supervised learning, which learns input–output functions.

  • These assumptions require researchers to assess statistical distributions and correlations before deciding whether ML is appropriate for a problem.
  • The relational hypothesis assumes a correlative relationship exists between model inputs and the system response.
  • The similarity hypothesis assumes that similar points in design space have similar properties or outputs.
  • Unsupervised learning identifies trends indicative of underlying behavior, whereas supervised learning learns a function f(x) = y encoding part of the PSPP relationship.

E. Unsupervised Machine Learning

Unsupervised learning identifies similarities, trends, and clusters in unlabeled AM data. Similarity metrics and k-means can partition manufacturing conditions and help decide which experiments are worth running.

  • E. Unsupervised Machine Learning: Unsupervised learning finds similarities or trends in unlabeled data and can identify manufacturing conditions with similar material properties.It is useful for visualizing high-dimensional datasets and screening irrelevant inputs.
  • E. Unsupervised Machine Learning: Vector distances provide a simple similarity measure that generalizes from a few manufacturing parameters to n design criteria.For points differing by δ in one coordinate, the ℓ2 distance reflects how similar their manufacturing conditions are.
  • E. Unsupervised Machine Learning: Under the similarity hypothesis, nearby manufacturing conditions are expected to yield similar measurements of a material property.The example treats points a, b, and c as similar when their distances are small.
  • E. Unsupervised Machine Learning: Similarity metrics can reduce the need to measure every possible design-space coordinate when selecting future AM experiments.They help determine whether an unmeasured test is sufficiently distinct to justify running it.
  • E. Unsupervised Machine Learning: Vector-distance clustering can produce partitions of material behavior analogous to process maps.These partitions provide a way to organize manufacturing conditions by similarity.
  • E. Unsupervised Machine Learning: K-means clustering partitions data into groups whose members have similar measurements or print parameters.The procedure repeatedly assigns points to the nearest centroid, recalculates centroids, and can expose input regimes associated with good or bad part quality.

F. Supervised Machine Learning

Supervised learning learns mappings from AM inputs to labeled outputs, using regression for continuous properties and classification for categories. Training minimizes a cost function, while error metrics and confusion matrices assess model performance.

  • F. Supervised Machine Learning: Supervised machine learning determines a functional relationship f(x) = y from previous measurements of labeled outputs.In AM, the inputs are manufacturing conditions and the outputs are measured material or process responses.
  • F. Supervised Machine Learning: Regression maps inputs to continuous values such as mechanical strength, whereas classification assigns inputs to output categories.The same AM inputs can support either task depending on the output being predicted.
  • F. Supervised Machine Learning: Least squares regression estimates coefficient vector β by minimizing the discrepancy between predicted and observed outputs.The closer Xβ − Y is to zero, the closer the model prediction is to the target relationship.
  • F. Supervised Machine Learning: A fitted functional relationship can predict a material property y′ for a new design-space point x′.The resulting approximation extracts additional information about the design space from existing measurements.
  • F. Supervised Machine Learning: Error metrics quantify model success and provide different information about prediction behavior, including mean prediction, outlier robustness, and uncertainty.The choice of metric affects what aspects of model performance are emphasized.
  • F. Supervised Machine Learning: MAE measures normalized absolute residuals, while MSE squares errors and therefore gives outliers greater influence.RMSE returns error in the same units as the predicted quantity, and MPE indicates directional prediction bias.
  • F. Supervised Machine Learning: Classification performance can be evaluated with a confusion matrix showing correct and incorrect class assignments.The matrix complements regression error metrics because classification outputs are non-numerical.

H. The Bias-Variance Tradeoff and Model Validation

Model validation balances fit complexity against generalization to unseen inputs. Training, test, validation splits, cross-validation, and comparisons across model types help detect overfitting, high bias, and high variance.

  • H. The Bias-Variance Tradeoff and Model Validation: A cost function quantifies the cost of a model parameterization, and training seeks parameters that minimize or maximize it.The loss function may also be called a reward function when the objective is maximization.
  • H. The Bias-Variance Tradeoff and Model Validation: Polynomial regression uses weights β_i as model parameters, while polynomial order is a model hyperparameter.Parameters describe a specific fitted relationship; hyperparameters define aspects of the model itself.
  • H. The Bias-Variance Tradeoff and Model Validation: The standard validation process divides data, estimates parameters on training data, tunes hyperparameters on test data, and validates the selected model.These steps repeat until the error estimate converges.
  • H. The Bias-Variance Tradeoff and Model Validation: Increasing model complexity can reduce training error while increasing out-of-sample error through overfitting.The bias-variance tradeoff balances representational bias against variance in unseen-data performance.
  • H. The Bias-Variance Tradeoff and Model Validation: Cross-validation trains models on subsets and evaluates them on held-out data to assess sensitivity to input selection and detect overfitting or selection bias.In k-fold validation, each fold serves once as the test set while the remaining k − 1 folds train the model.
  • H. The Bias-Variance Tradeoff and Model Validation: Leave-one-out cross-validation uses k = n and is especially useful for assessing the effect of outliers on model performance.Each model trains on all points except one and tests on the remaining point.
  • H. The Bias-Variance Tradeoff and Model Validation: LOCO cross-validation was introduced for materials datasets with clustered composition or property distributions.Its purpose is to expose distributional problems that ordinary validation may miss in materials and AM datasets.
  • H. The Bias-Variance Tradeoff and Model Validation: Comparing multiple algorithms is useful because different models make different assumptions and may perform differently on the same dataset.The confusion matrix example illustrates classification outcomes by class, including correct diagonal entries and off-diagonal errors.

I. Comparison Across Machine Learning Approaches

AM researchers should evaluate candidate algorithms systematically rather than assume one model is best. The recommended workflow combines data assessment, preprocessing, tuning, cross-validation, multiple metrics, and algorithm comparison using available software tools.

  • I. Comparison Across Machine Learning Approaches: Finding the best parameterization of one model does not guarantee the best solution across all reasonable algorithms.Different approaches should be evaluated for the same AM problem before selecting a model.
  • I. Comparison Across Machine Learning Approaches: PCA and KRR embody different assumptions about input-output relationships, so both may be tested when linearity is uncertain.PCA assumes linearity, whereas KRR does not.
  • I. Comparison Across Machine Learning Approaches: Model selection for AM includes checking statistical correlations, preprocessing and featurizing data, tuning parameters, and applying cross-validation.These steps precede comparison of error metrics across algorithms.
  • I. Comparison Across Machine Learning Approaches: Researchers should compare error metrics across several algorithms and select the best performer for the application.Multiple metrics matter because algorithms handle outliers and statistical correlations differently.
  • I. Comparison Across Machine Learning Approaches: Tools for comparing algorithms can improve overall performance by enabling selection from multiple reasonable approaches.The review identifies comparison capabilities in commonly used machine-learning packages.
  • I. Comparison Across Machine Learning Approaches: Commonly used software includes machine-learning, statistical-analysis, and computer-vision toolboxes, with both open-source and commercial options represented.Table IV highlights their relevance to AM synthesis optimization.

III. CURRENT ICME TOOLS ARE WELL EQUIPPED TO INTEGRATE WITH AN ML FRAMEWORK

Machine learning can connect existing physics-based materials models and databases to metals AM development, supporting classification, alloy exploration, and composition-based property prediction. The review emphasizes that AM-specific optimization remains constrained by incomplete data and the difficulty of producing combinatorial feedstock.

  • III. CURRENT ICME TOOLS ARE WELL EQUIPPED TO INTEGRATE WITH AN ML FRAMEWORK: The review discusses connecting machine learning approaches with physics-based analyses, characterizations, and simulations to automate AM process-structure-property-performance knowledge generation.It presents selected approaches demonstrated in materials science or with clear, immediate AM applications rather than an exhaustive review.
  • Alloy Design and Feedstock Selection: Alloy composition affects AM physics from energy-source interactions and feedstock density to melt conduction, solidification kinetics, vaporization, stoichiometry, microstructure, and final quality.These effects also vary with feedstock type, including wire versus powder.
  • Classification: ROC curves compare true-positive and false-positive rates for binary classifiers, while AUC summarizes classification performance.The example classifies material-quality regimes such as high density or low density from AM machine inputs using literature-mined data.
  • Classification: Naïve Bayes significantly outperforms the other two algorithms in the example ROC comparison and is therefore the best machine-learning choice for that task.A perfect classifier has AUC = 1.
  • Database mining: Database mining can predict stable crystal structures and composition-dependent material properties, and established search algorithms can extend these capabilities to AM alloy cases.Examples include ICSD, the Linus Pauling files, AFLOWLib, and the Materials Project.
  • Limitations: Database mining is limited to previously measured or calculated properties, while combinatorial AM feedstock studies are hindered by the difficulty and expense of producing feedstock.Models and optimization targets for extracting AM candidate alloys are still being developed and verified.
  • Genetic algorithms: 15,503 alloy combinations arise when Ti is combined with Al, V, Zr, Cr, and Hf in 1 wt. % steps up to 15% total alloying.This illustrates why explicitly creating every candidate can be undesirable for AM feedstock exploration.
  • Genetic algorithms: Genetic algorithms evaluate candidate-alloy fitness with physics-based models, then use selection, crossover, and mutation to approach a target property without explicitly modeling every combination.The evaluation function may use DFT, ThermoCalc, or additive-specific composition-property models.

2. Design of Experiments

Machine-learning-based design of experiments performs parametric analysis iteratively, using each result to select the next experiment. In AM, this framework can explore or optimize process-structure-property relationships while reducing the number of experiments needed.

  • 2. Design of Experiments: Design of Experiments maps independent variables to dependent parameters, and machine learning can reduce experiments needed to characterize AM process-property relationships.AM studies vary parameters such as laser energy, speed, build direction, composition, and layer height.
  • 2. Design of Experiments: Information-gain experiments reveal previously unobserved process-structure-property relationships, supporting theory development in science and optimality criteria such as strength or porosity in engineering.Both disciplines vary inputs and measure responses until sufficient information is obtained.
  • 2. Design of Experiments: Traditional DOX predetermines and completes tasks before analysis, whereas machine-learning DOX analyzes each result and selects the next task using a statistical metric.The adaptive approach updates the experiment sequence after every individual task.
  • 2. Design of Experiments: ML-based DOX begins by identifying process-structure-property inputs and outputs, defining the design space, generating an initial dataset, and specifying a response function.An initial uniform sample can provide the starting information, followed by a regression model of inputs and material properties.
  • Decision trees: Decision trees choose feature partitions that minimize Gini impurity, which can extend to arbitrary classes and support classification or regression.In the example, splitting at x = 0.2 gives zero impurity on both sides and perfect classification.
  • Decision trees: Random forests address decision-tree overfitting by training many trees on bootstrapped observations and, optionally, subsets of features.The ensemble can also provide uncertainty metrics for predictions, depending on the application.
  • 2. Design of Experiments: Random forests train many decision-tree regressors on data subsets, predict properties at new design points, and support experiment selection through improvement or uncertainty response functions.MEI or MLI selects experiments for maximizing a target, while MU favors the most uncertain point for exploration.

3. Topology Optimization and Generative Design

Topology optimization connects material, process, and geometry decisions to optimize AM part performance, while machine learning accelerates design exploration and enables generative approaches.

  • Topology Optimization: Topology optimization designs geometry-material-performance relationships to maximize performance, minimize weight, or reduce manufacturing costs.AM-specific constraints include overhangs, support structures, and part density.
  • Machine Learning for Materials and Design: Random forests and sequential learning identify optimal material candidates more quickly than random guessing across multiple materials datasets.The random forest approach was also compared with COMBO as a sequential learning tool.
  • Topology Optimization: Topology optimization is a local method, so global design optimization generally requires statistical analysis of many topology-optimization simulations.This creates a computational burden when exploring broad design spaces.
  • CNN-Based Design: CNN-based one-shot tools predict optimized two- and three-dimensional structures directly from initial geometry and loading conditions, with accuracy similar to traditional optimization.The models use optimized topologies generated by traditional topology optimization for training and validation.
  • Generative Design: GANs use competing generator and discriminator networks to produce increasingly realistic optimized topologies and designs.Combined CNN-GAN methods can upscale coarse topology results to higher resolution with substantially lower computational time.
  • Generative Design: A combined CNN-GAN produced high-resolution topologies within 3% of expected ground-truth pixel values and in 0.06% of traditional topology-optimization time.The CNN first predicts a low-resolution topology, which the GAN then converts into a high-resolution design.

B. Machine Learning Assisted Modeling of Additive Manufacturing

Machine learning can support computational design of additive-manufacturing processes by modeling input-output relationships and revealing trends in high-dimensional design spaces.

  • Scope: The review examines machine learning algorithms as tools for computational design of additive-manufacturing process developments.The design space of AM experiments is often vast, making these methods relevant to process-centric engineering.
  • Modeling Process Relationships: Machine learning provides alternative numerical methods for calculating y = f(x) from experimental measurements or previously run simulations.These models can replace or complement direct numerical calculations in process development.
  • High-Dimensional Analysis: Machine learning also helps visualize trends in high-dimensional spaces, enabling researchers to study complex relationships and ask deeper questions.This addresses process-centric engineering objectives alongside numerical prediction.

1. Machine Learning as Numerical Methods for Modeling

Machine learning reduces the computational burden of additive-manufacturing modeling by selecting useful simulations, removing irrelevant dimensions, or approximating numerical methods.

  • Existing Numerical Methods: Finite-element and finite-volume methods remain established numerical approaches for modeling feedstock, heat-source, and melt-pool dynamics in AM.AM-specific adaptations to finite-element methods have also been developed.
  • Computational Efficiency: Machine learning can make AM modeling more efficient by determining which models not to run, reducing dimensionality, or replacing explicit numerical methods.The reviewed alternatives include finite-element and cellular-automata calculations.
  • Dimensionality Reduction: Dimensionality-reduction methods identify relevant parameters or physics, enabling reduced models that achieve the same result with lower computational burden.Materials-science examples show dimensionality reduction relating local atomic environments to mesoscale properties.
  • Design of Experiments: Random forests can guide design of experiments by predicting outputs for unobserved points and focusing attention on influential parameters.The method trains on previously completed experiments represented as rows of a data matrix.
  • Feature Screening: Kamath identified laser speed and power as the most important of four inputs for predicting melt-pool depth and shape.This screening allowed more expensive modeling to focus on the parameters most relevant to the process response.
  • Matrix Factorization: Principal Component Analysis represents high-dimensional data through components ordered by variance, allowing regression with less data and minimal information loss.In AM, the resulting vectors can serve as regression inputs for predicting material properties.

2. Machine Learning for Visualizing Trends in the Design Space

Machine learning helps visualize and model high-dimensional AM design spaces, where conventional process maps are limited and expensive simulations cannot cover all manufacturing conditions.

  • Design-space visualization: Traditional process maps show relationships among machine inputs and properties but generally interpret only a few parameters at once.Extending process maps to many process variables becomes difficult to visualize.
  • Design-space visualization: tSNE projects high-dimensional AM data onto two dimensions while preserving similarity relationships between nearby points.This provides an interpretable visualization of how machine inputs and material behavior are distributed.
  • Design-space visualization: tSNE can cluster samples using many manufacturing inputs simultaneously, allowing engineers to examine processing conditions associated with good or bad material properties.A comparison of process maps and tSNE is shown in Figure 10.
  • Modeling AM processes: AM model generation is difficult because understanding how process parameters affect material properties and performance requires incorporating substantial physics.Traditional thermal-history, residual-stress, and thermal-signature models have been applied, but full-physics simulations can be computationally expensive.
  • Modeling AM processes: Surrogate models estimate high-cost simulation results from previously run simulations and interpolate outcomes at new design-space coordinates.Their accuracy depends on the number and distribution of previous simulations.
  • Modeling AM processes: Tapia et al. used a Gaussian process model to predict melt pool depth from laser powder bed fusion simulation inputs and accurately predicted an unobserved condition.The model predicted both high-fidelity simulation outcomes and experimental measurements, supporting faster exploration of manufacturing states.
  • Modeling AM processes: Gaussian process response surfaces can identify design-space regions that provide specific material responses, supporting inverse design and additive manufacturing optimization.Tapia et al. visualized melt-pool depth as a function of laser power and speed.
  • Modeling AM processes: Machine learning may combine multiobjective optimization with real-time computer vision to observe, characterize, and control multiple AM process aspects simultaneously.The proposed targets include energy density, melt-pool shape, heat transfer, grain growth, and final part properties.

C. Process Monitoring and Control

AM process monitoring and control require rapid analysis of diverse, multiscale signals. Machine learning and computer vision can extract features, detect deviations, and support feedback across sensors.

  • Monitoring objectives: AM monitoring spans many sensor types, phenomena, and length and time scales, making data collection and control difficult.Practical monitoring requires compromises between collection rates and resolutions.
  • Monitoring objectives: In situ monitoring seeks to control microstructure development and prevent defect formation during AM.Experiments also identify process conditions associated with desired outcomes or defect avoidance.
  • Computer vision and signal processing: Computer vision can monitor multiple sensor types and signal deviations, such as temperature spikes or sharp image-intensity changes.These deviations can indicate departures from desired processing conditions.
  • Computer vision and signal processing: Image filters reduce noise and extract features from AM time-series and image data, including edges and object boundaries.Median, conservative, and Gaussian filters address different noise conditions, while edge detectors identify local intensity transitions.
  • Computer vision and signal processing: Template matching identifies AM patterns such as abnormal melt-pool morphologies, unmelted powder particles, and denudation near the melt zone.SIFT and SURF are feature-identification methods used for these pattern-recognition tasks.
  • Computer vision and signal processing: If an in situ feature matches a defect-indicative feature in a dictionary, a defect likely formed during manufacturing.Machine learning can leverage correlations from previous measurements, observations, and responses to fill monitoring gaps.
  • Neural networks: Neural-network activation functions introduce nonlinearity, while increasing network depth and width enables regression and classification.The sparsity of materials data and complexity of process–structure–property relationships require attention to model tradeoffs and activation functions.

2. Featurization of Qualitative Image Data

Machine-learning computer vision can automate qualitative image characterization in AM, reducing the burden of measuring microstructure, powder, and material classes.

  • Automated image characterization: Computer vision can automate microstructure classification during AM parametric analysis.This targets measurements that otherwise require tedious analysis of strength, roughness, microstructure, or defect-density data.
  • Automated image characterization: AM image-analysis algorithms generally require tailoring for AM microstructure data rather than functioning out of the box.Methods proven on non-AM microstructure datasets may nevertheless be extensible to AM datasets.
  • Applications: SIFT, k-means clustering, and an SVM classified synthetic AM powder images into particle sizes with 89% overall classification accuracy.Particle-size distributions affect energy absorption and part metrology across the additive process.
  • Applications: Miyazaki et al. used image filters and a random forest to segment selective-laser-melted Ti-6Al-4V images into α- and β-phase regions.The method automatically calculated phase area fractions and measures related to grain morphology and distribution.
  • Applications: Machine-learning ensembles have been used to classify microstructures as dendritic or non-dendritic, while broader efforts automate materials classification from metallographs.These approaches are intended to accelerate materials characterization and development.

IV. LEARNING FROM THE PAST: MOVING TOWARDS DATABASE-DRIVEN DESIGN OF ADDITIVE TECHNOLOGIES

Database-driven design can expand the AM design space by integrating process, structure, property, and performance data. Openly shared datasets and infrastructure are identified as important enablers.

  • Database-driven AM design: AM research generates data from parametric analysis, computational modeling, and in situ monitoring, and integrating these data expands the design space explored.The integrated datasets can span multiple parameters, physical processes, printers, length scales, and time scales.
  • Lessons from materials science: Materials-science databases have long supported composition–property searches, high-throughput experimentation, and computational materials design.Examples include crystal-structure databases, high-throughput characterization libraries, and DFT-based workflows.
  • Lessons from materials science: Machine learning has been used since at least 2005 to address high-throughput DFT obstacles such as computational expense, limited transferability, and large design spaces.Neural networks, genetic algorithms, and principal component analysis can approximate properties of systems too large for direct DFT calculation.
  • Open data infrastructure: Open-source databases such as the Materials Project combine experimental and computational materials information with automated DFT workflows.Accessible databases form a primary component of the Materials Genome Initiative roadmap.
  • Open data infrastructure: Open sharing of AM microstructure image databases can support segmentation and identification algorithms tailored to materials and AM-specific problems.The review argues that AM should develop infrastructure for open data sharing comparable to other materials fields.

V. CONCLUSIONS

Machine learning is presented as a complementary tool for addressing metals AM’s complex, interdisciplinary design and manufacturing challenges. The review identifies applications spanning coupled physics-statistics models, integrated materials and process design, automated control, and data infrastructure, while emphasizing data quality, sensing, standards, and physics integration as remaining needs.

  • Conclusions: Machine learning is not an end-to-end solution: its use must be matched to suitable problems and algorithm classes, alongside experiments and physics-based modeling.The review frames ML as a complementary technology for addressing unresolved process-structure-property-performance challenges in AM.
  • Coupled Physics-Statistics Models: Machine learning can provide computationally accessible surrogate models for metals AM systems governed by coupled optical, fluid, mechanical, thermal, and kinetic physics.
  • Materials Design: Machine learning can jointly optimize material choices and manufacturing conditions, including composition, phase kinetics, energy density, deposition rate, and feedstock supply.In situ control may be part of process optimization.
  • Automated Process Control: Intelligent feedback and control can integrate multiple process signals and optimize multiple objectives, supporting tighter control of additive manufacturing.Signal processing, computer vision, sensors, and measurement techniques contribute to these control models.
  • Data Sharing Infrastructure: Shared AM-data infrastructure could accelerate data-driven materials design, but standardizing AM data types is necessary for useful and meaningful sharing.
  • Further Developments: AM machine-learning development remains constrained by data curation, standards, sensor capabilities, and the need to build algorithms around validated physics-based models.Fast, microscale melt-pool imaging and temperature-history measurements are identified as important sensor-development needs.
Loading 2005.05235v1…