PaperlayerView exact revision on arXiv

Source-linked AI summary

Machine Learning for Material Characterization with an Application for Predicting Mechanical Properties

Anke Stoll, Peter Benner

arXiv:2010.06010v1cs.LG

TL;DR

Material-property data are growing, while obtaining measurements through experiments can be expensive and time-consuming. The paper reviews machine learning for metallic characterization and applies it to small punch test data. It reports correlations supporting prediction of tensile strength for five steels across nine heats, with no further tensile tests needed once the model is established.

  • Problem

    Material-property measurements can be expensive and time-consuming, and growing data volumes create challenges for processing and analysis.

    Source: §1 · p. 2

  • Method

    The paper surveys machine-learning applications in metallic material characterization and applies machine-learning models to small punch test data to predict tensile properties.

    Source: §1 · p. 2

  • Results

    Small punch test curves correlate with R_m and can predict it for five different steels (nine heats); once the model is established, further tensile tests are not needed.

    Source: §3.4 · p. 14

  • Takeaways & Limitations

    The reported results support using simple small punch tests with machine learning to predict tensile properties such as ultimate tensile strength.

    Source: §3.4 · p. 14, §4 · p. 15

    The available data are very limited, so generalization of the linear model and random forest to other materials, temperatures, and geometries requires critical evaluation.

    Source: §3.4 · p. 14

Abstract

from arXiv · show

Currently, the growth of material data from experiments and simulations is expanding beyond processable amounts. This makes the development of new data-driven methods for the discovery of patterns among multiple lengthscales and time-scales and structure-property relationships essential. These data-driven approaches show enormous promise within materials science. The following review covers machine learning applications for metallic material characterization. Many parameters associated with the processing and the structure of materials affect the properties and the performance of manufactured components. Thus, this study is an attempt to investigate the usefulness of machine learning methods for material property prediction. Material characteristics such as strength, toughness, hardness, brittleness or ductility are relevant to categorize a material or component according to their quality. In industry, material tests like tensile tests, compression tests or creep tests are often time consuming and expensive to perform. Therefore, the application of machine learning approaches is considered helpful for an easier generation of material property information. This study also gives an application of machine learning methods on small punch test data for the determination of the property ultimate tensile strength for various materials. A strong correlation between small punch test data and tensile test data was found which ultimately allows to replace more costly tests by simple and fast tests in combination with machine learning.

1 INTRODUCTION

Materials science faces growing data volumes and high dimensionality, while material databases can remain small because experiments and simulations are costly. This paper reviews machine learning for materials characterization and examines whether small punch test data can predict tensile properties.

2.1 Overview - machine learning for materials science

  • The overview situates data-driven materials research within a progression from experiments and equations to computational simulations and data-driven discovery of patterns.The paper also surveys machine-learning applications including property prediction, candidate identification, and experimental design.Source: §1 · p. 2, §2.1 · p. 3

1 INTRODUCTION

  • Growing data volumes and attributes create high-dimensional data whose associated learning challenges include unfavorable effects on algorithm behavior and performance.Materials databases may also be small because data acquisition through experiments or simulations is expensive and time-consuming.Source: §1 · p. 2
  • The paper surveys machine learning for metallic material characterization and investigates whether tensile properties can be predicted from small punch test data.It also reviews machine learning approaches based on small punch test data.Source: §1 · p. 2

2.1 Overview - machine learning for materials science

  • Machine learning can analyze process–structure–property–performance relationships, although there is no general agreement on how these elements are interconnected.Stated aims include learning intrinsic interrelations and high-throughput identification and quantification of material properties.Source: §2.1 · p. 2

Open problem - interpretability

Machine-learning models can be difficult to interpret, limiting the understanding and knowledge gained from their use. The section discusses model and data-reduction choices that can improve interpretability, while noting that accuracy may be traded for explainability.

Open problem - interpretability

  • Complex machine-learning models are often treated as black boxes that are difficult for humans to understand.This is identified as a criticism because their use may not yield novel understanding or knowledge.Source: Open problem - interpretability · p. 3
  • Scientific models should support clear causal relations between broadly defined material structure and properties while using few parameters and adhering to physical laws.The section presents interpretable models as useful for translating results into knowledge other scientists can use.Source: Open problem - interpretability · p. 3
  • PCA reduces dimensionality by extracting orthogonal directions of greatest variance, but its components may not be physically simple to interpret.The section also describes feature selection and regularization as ways to reduce the feature space.Source: Open problem - interpretability · p. 3
  • Regressions and decision trees offer more direct interpretability than complex neural networks, and accuracy may sometimes be traded for explainability.Feature visualization and attribution are also cited as ways to improve understanding of black-box models.Source: Open problem - interpretability · p. 3

Open problem - small data

Materials characterization often has limited training data because experiments and simulations are costly, time-consuming, and constrained by testing availability. The section reviews approaches to small-data modeling and identifies database accessibility and publication practices as additional concerns.

Open problem - small data

  • Experimental and simulated data are costly to generate, and models fitted to extremely small training sets may not replicate machine learning successes in other fields.Small datasets can still be used for simple fitting in low-dimensional settings.Source: Open problem - small data · p. 3
  • Small-data approaches include restricting the modeled material space, adding crude property estimates as features, or integrating prior domain knowledge and physical constraints.The section also describes partial labeling, visualization, and monotonization as ways to incorporate domain knowledge.Source: Open problem - small data · pp. 3–4

Existing databases

  • Unpublished negative results and data left out of publications can bias machine-learning datasets and lead researchers to repeat work to build balanced data.The section identifies this as publication bias and wasted resources.Source: Existing databases · p. 4
  • Materials-property databases support materials informatics, but few are ready for informatics software because they lack uniform formats or application programming interfaces.The section names several existing databases and describes efforts to make materials data more accessible.Source: Existing databases · p. 4

Materials informatics

Materials informatics encompasses organizing and analyzing materials data, with iterative modeling proposed as a route to simple, interpretable models. For metallic characterization, machine learning uses experimental or simulated data to relate material properties to their characteristics.

Materials informatics

  • A materials-informatics workflow proceeds from primary features through exploratory modeling and refinement to final training and deployment, aiming for a simple, interpretable model.The section also emphasizes data standardization, organization, access, and knowledge discovery.Source: Materials informatics · p. 4

2.2 Machine learning for metallic material characterization

  • Machine learning for material-property prediction maps relationships between properties and related characteristics from experimental or simulated data.Mechanical properties are linked to processing parameters and resulting microstructures, and simulations often describe mechanical behavior with constitutive equations.Source: §2.2 · p. 5
  • Machine learning can help estimate material parameters when elementary physical ideas cannot reliably estimate them, using experimental or simulation data.Research on macroscopic performance often focuses on relationships between material properties and microstructure.Source: §2.2 · p. 5
  • The section reviews efforts to enhance destructive metal tests, including tensile, fatigue, hardness, and impact tests, with machine-learning methods.Source: §2.2 · p. 5

Corrosion

Machine learning is applied to corrosion detection, classification, and forecasting using image data and material or environmental variables. Reported approaches range from CNNs and SVMs to hybrid GA-SVR models.

Corrosion

  • Corrosion monitoring matters for infrastructure health and life-cycle cost, and ML approaches may reduce inspection time and cost while aiding human inspectors [79–81].Source: Corrosion · p. 5
  • CNNs classify image regions as corroded or not and outperform texture- and color-based approaches; pretrained networks can help avoid overfitting on small datasets [81].The input is a material image and region of interest, evaluated with a sliding window.Source: Corrosion · p. 5
  • Bastian et al. classify corrosion images into four levels: none, low, medium, or high.Source: Corrosion · p. 5
  • A GA-SVR hybrid forecasts atmospheric corrosion depth in zinc and steel, with the genetic algorithm selecting SVR hyperparameters.Inputs include temperature, time of wetness, exposure time, sulphur dioxide, and chloride concentrations.Source: Corrosion · p. 5
  • For pipe corrosion, SVMs classify image blocks using surface-texture features, while SVR predicts 3C steel corrosion rate from 46 samples with very small error [84, 85].The pipe-classification SVM is optimized by differential flower pollination.Source: Corrosion · p. 5
  • ANNs and linear-kernel SVMs outperform kNN and classification trees for pitting-corrosion detection in 316L stainless steel, without requiring surface analysis.The models use chloride concentration, pH, and temperature as inputs.Source: Corrosion · pp. 5–6

Fatigue

The section surveys ML applications to fatigue-strength and fatigue-life prediction from material, processing, and test data. Reported approaches include deterministic and model-based learning, ensembles, and hybrid GA-SVM methods.

Fatigue

  • A model-based ANN for 40 low-carbon steels incorporates microstructures and stress-strain curves, allowing prior structure-property knowledge and uncertainty such as fatigue-life scatter to be considered.The contrasted deterministic models include multivariate linear regression and ANNs using composition, processing, inclusion sizes, and fatigue strength.Source: Fatigue · p. 6
  • Ensemble methods and methods individualized for different material types showed the most success in predicting steel fatigue strength across the approaches reviewed.The study considered regression, decision trees, SVMs, and ANNs using composition and processing-related inputs.Source: Fatigue · p. 6
  • A combined genetic algorithm and SVM predicts P91 steel base-metal fatigue life from fatigue-test hold time with high accuracy [88].Source: Fatigue · p. 6
  • An ANN radial-basis-function model predicts fatigue life of steel reinforcing bars using maximum tensile strain and pressure ratio.Source: Fatigue · p. 6

Creep

The section describes ML for creep and creep-rupture prediction, where material behavior depends on interacting composition, processing, and microstructure variables. Reported results range from insufficient model performance to higher accuracy with selected features.

Creep

  • Creep-rupture forecasting and useful-life estimation address high-temperature power-system risks, while ML is presented as an alternative to long, expensive experiments for behavior shaped by multidimensional dependencies [90, 91].Source: Creep · p. 6
  • Creep-rupture life and stress prediction for austenitic stainless steels used ANN inputs including test conditions, composition, heat treatment, quenching, and grain size [93].Solution-treatment time was available in only a very limited number of cases.Source: Creep · p. 6
  • General Regression Neural Networks, ANNs, and Gaussian Processes had insufficient overall performance for creep-rupture forecasting from experimental data.Source: Creep · p. 6
  • Random forest achieved the highest accuracy for predicting Larson-Miller parameters among five models, with top-feature counts ranging from 5 to 21.The study selected relevant features from 466 candidates using optimization approaches.Source: Creep · p. 6
  • For rupture and creep-rupture stress prediction in 9%Cr steels, a multilayer perceptron used composition, heat-treatment information, and component geometry [91].Source: Creep · p. 6

Flow behaviour and work hardening

The section reviews ML for flow behavior and work-hardening characterization, including hot-deformation prediction and indentation-based parameter identification. It also contrasts conventional tests and mathematical models with data-driven approaches.

Flow behaviour and work hardening

  • Tensile tests provide uniaxial stress-strain behavior, while compression testing is complicated by friction and buckling and bend or torsion testing by stress and strain gradients [95, 96].Hardness correlates well with flow stress.Source: Flow behaviour and work hardening · p. 7
  • A feed-forward ANN for 42CrMo steel predicts flow-stress level and strain-hardening and flow-softening stages from temperature, log strain rate, and strain, with low absolute relative errors on training and test data.Source: Flow behaviour and work hardening · p. 7
  • For flow-stress prediction, ANN and SVR approaches using strain, strain rate, and temperature produced more accurate results than conventional models in the cited studies.The comparisons included Johnson-Cook, modified Zerilli-Armstrong, and modified Arrhenius models.Source: Flow behaviour and work hardening · p. 7
  • An indentation protocol using imprint shape with orthogonal decomposition and manifold learning predicts strain-hardening exponent and yield stress in steel and an aluminum alloy.Source: Flow behaviour and work hardening · p. 7
  • An ANN uses strain and temperature to predict ultimate strength, yield strength, elongation, and strain-hardening coefficient in Inconel 718 at elevated temperatures.Source: Flow behaviour and work hardening · p. 7

Tensile properties

Machine learning models have been applied to predict tensile properties from composition, processing, and test data, while small punch testing offers a small-specimen route to material characterization. Small punch analysis spans empirical correlations and ML-assisted identification, with inverse solutions and test conditions imposing important limitations.

Tensile properties

  • Yield strength, UTS, and elongation were modeled from steel composition and processing; reported approaches linked strengthening mechanisms to ductility tradeoffs and used multi-objective optimization.Hot-ductility models included a cooling condition unlike commercial continuous casting, and one study reported limited data and a single hidden layer [117, 105].Source: Tensile properties · pp. 7–9
  • ANNs predicted several mechanical properties from material and process inputs, while ensembles provided more reliable predictions than individual models for weld metals.ANNs also predicted casting UTS [104], and acoustic-emission parameters were used to predict composite and tensile-specimen strength [116].Source: Tensile properties · pp. 7–8
  • LWR, least-squares SVMs, and random forests were used for tensile strength, elastic modulus, yield stress, and steel-design properties [111–113, 92, 103].For cast alloys, models used micrograph information directly, with ANNs performing best for UTS and yield strength.Source: Tensile properties · p. 8

2.3 Machine learning for the small punch test

  • Small-specimen tests are attractive when material volume is limited, and SPT has been used to estimate flow and other properties, with correlations to conventional tensile testing serving as a prerequisite [8, 121–123, 127, 128].Reported examples include linear correlations between SPT maximum force and tensile strength in 9% chromium steel [134].Source: §2.3 · p. 9
  • SPT inverse identification is non-unique and can be affected by measurement and modeling issues; small samples may not represent bulk material, and response depends on test setup [133, 140, 141, 163].The sample thickness is described as 5–8 times the mean grain size, and relevant setup variables include specimen geometry, thickness, speed, ball diameter, and clamping force.Source: §2.3 · pp. 9–10
  • ML-assisted SPT characterization includes FEM-trained ANNs for parameter identification and manifold learning for matching simulated and experimental imprint shapes [140, 133].Other approaches use ANNs to identify damage-model parameters [166, 167] or image processing of Erichsen tests to estimate anisotropy, though results were environment-biased [168].Source: §2.3 · p. 10

3 AN APPLICATION OF ML IN MATERIAL PROPERTY PREDICTION

This application tests whether machine learning can predict tensile strength from small punch test data, using experiments across multiple steels and conditions. The authors report that SPT curves correlate with UTS and can support its prediction, while noting limits to model generalization.

3 AN APPLICATION OF ML IN MATERIAL PROPERTY PREDICTION

  • The application investigates whether machine learning can determine tensile properties from SPT data, proposing two ML models alongside empirical correlations.The empirical approach requires a geometry- and material-dependent correlation factor.Source: §3 · pp. 10–11

3.1 Experiments

  • Experiments used SPT data from three ferritic-martensitic Cr-steels and two bainitic reactor pressure vessel steels, with selected materials tested at multiple temperatures.The dataset spans nine materials and heats; the P91 curve at −177 °C represents brittle failure, while the other shown test curves represent ductile failure.Source: §3.1 · p. 11, Figure 2 · p. 12, Figure 1 · p. 12

3.2 Preprocessing

  • Each SPT curve was sampled at 44 μm intervals up to 1.5 mm, yielding 151 displacement features plus temperature; PCA addressed feature collinearity, with six components explaining 99% of variability.The pairs plots compare selected raw features and temperature with the first five principal components and UTS.Source: §3.2 · pp. 12–13, Figure 3 · p. 13

3.3 Machine learning models

  • With limited data, the study compares PCA-based linear regression with random forest on original SPT data, using RMSE and 10-fold cross-validation.The empirical approach is also evaluated; Figure 4 plots empirical, LM-based, and RF-based UTS against experimental UTS.Source: §3.3 · p. 13, Table 3 · p. 14, Figure 4 · p. 14

3.4 Application conclusions

  • LM combines multiple principal components rather than one SPT force value and may explain its slightly higher accuracy, but PCA makes the resulting model less intuitive to interpret.The authors caution that generalization across other materials, temperatures, and geometries needs critical evaluation; they also flag RF use outside 500–1000 MPa as potentially unsuitable.Source: §3.4 · p. 14
  • SPT curves predict UTS across five steels (nine heats); the LM approach is slightly more accurate than the empirical approach, and established models could remove further tensile tests.The authors also state that broader data may be needed to re-establish the models.Source: §3.4 · p. 14

4 DISCUSSION AND CONCLUSIONS

The survey presents machine learning as a way to find material structure–property relationships that are difficult to capture across scales and timescales. Its application example reports good UTS prediction from small punch test data, while emphasizing data and model limitations.

4 DISCUSSION AND CONCLUSIONS

  • Machine learning can help identify material structure–property relationships that are often nonlinear and difficult to discover across multiple length and timescales.The survey describes data-mining as important for recognizing correlations in experimental and simulated data.Source: §4 · pp. 14–15
  • Limited material data can compromise model learning; models trained on small accurate datasets may not generalize beyond their input space, while broader coverage may reduce precision.The discussion also identifies missing failure data as an obstacle to precise ML models.Source: §4 · p. 15
  • Artificial neural networks have modeled complex materials behavior, but conventional networks face overfitting and generalization concerns and can be difficult to stabilize.The paper also notes their use for predicting microstructures and phases after heat treatment or deformation.Source: §4 · p. 15
  • No single model is best across applications; the appropriate choice depends on the specific application and data situation.The discussion notes that hybrid and ensemble methods can also work well.Source: §4 · p. 15
  • Simple machine learning approaches, including linear regression and random forests, provide good UTS predictions from small punch test data, even with a very small database.The authors conclude that these results confirm the benefit of simple ML for predicting mechanical properties from simple material tests.Source: §4 · p. 15
  • Small punch testing is inexpensive and simple, but specimen size may not represent bulk material, sample-size effects influence properties, and results are test-parameter-sensitive.The paper frames SPT-based prediction as an application for material characterization, including failure analysis and remaining-life assessment.Source: §4 · p. 15

Related papers

Loading 2010.06010v1…