Search

The application of machine learning (ML) to corrosion research has become an important trend in corrosion science in recent years. In this paper, the feature extraction method for corrosion data and the ML algorithms commonly used (including artificial neural networks, support vector machines, ensemble learning and other widely used algorithms) in corrosion field is introduced. Then, the characteristics of different algorithms and their application scenarios in the corrosion prediction are summarized. Finally, the development trend of ML in material corrosion field is prospected. [ABSTRACT FROM AUTHOR]

Data mining was introduced and 178 sets of alternating current (AC) corrosion data were collected from published data and our research group. The AC current density (JAC), direct current (DC), current density (JDC), the ratio of AC/DC current density (JAC/JDC), and cathodic protection potential (IR‐free potential, EIR‐free) were defined as the input feature and the priority of feature subset was ranked as (JAC, JDC) > (JDC, JAC/JDC) > (JAC, EIR‐free) >JAC > (JAC/JDC, EIR‐free) > JAC/JDC. Then, based on the different feature subsets, the AC corrosion rate model was established by the random forest algorithm, and the generalization ability of the model was verified on the test data respectively. Finally, the mapping relationship between AC, DC parameters, and corrosion rates was presented and the limiting values of (JAC, JDC) and (JDC, JAC/JDC) were given. [ABSTRACT FROM AUTHOR]

Coupon tests on the pipeline subjected to metro dynamic DC interference were conducted. The pipe‐to‐soil from 78 sites including off‐potentials (Eoff), on‐potentials (Eon), and corrosion rate data were obtained, and then the dynamic fluctuation parameters of pipe‐to‐soil potentials were analyzed statistically. Further, the Pearson correlation between dynamic statistical parameters of pipe‐to‐soil potentials and corrosion rates was analyzed and showed that the statistical parameter of "roff ⩾ −0.85 V" and "mean[Eoff,cmax]" have a greater correlation with corrosion rate than others. Based on the statistical parameters of Eoff and the machine learning method, the optimal feature subset of the corrosion rate model was selected; meanwhile, it is proved that the negative shift of Eoff plays an important role in the prediction of corrosion rate. Finally, the corrosion rate model for buried steel pipeline under dynamic DC interference was established by random forest (rfr) method on 66 training data, then generalization performance of the model was verified on 12 test data. The results show that mean absolute error and R2 are 0.0358 mm/a and 0.56, respectively, on test data. [ABSTRACT FROM AUTHOR]

Data provides a foundation for machine learning, which has accelerated data-driven materials design. The scientific literature contains a large amount of high-quality, reliable data, and automatically extracting data from the literature continues to be a challenge. We propose a natural language processing pipeline to capture both chemical composition and property data that allows analysis and prediction of superalloys. Within 3 h, 2531 records with both composition and property are extracted from 14,425 articles, covering γ′ solvus temperature, density, solidus, and liquidus temperatures. A data-driven model for γ′ solvus temperature is built to predict unexplored Co-based superalloys with high γ′ solvus temperatures within a relative error of 0.81%. We test the predictions via synthesis and characterization of three alloys. A web-based toolkit as an online open-source platform is provided and expected to serve as the basis for a general method to search for targeted materials using data extracted from the literature. [ABSTRACT FROM AUTHOR]

Simulating the frictional properties of complex interfaces is computational resource consuming. In this paper, we propose a density functional theory (DFT) calculation combined machine learning (ML) strategy to investigate the sliding potential energy corrugation between geometrical corrugated graphene (Gr) sheets. By the aid of few DFT calculations and geometrical descriptors Σr−n (n = 1, 2, 6, 12), the trained ML models can accurately predict the sliding potential evolutions of Gr/Pt and Gr/Re systems. To be specific, based on DFT calculations of sliding along [110] direction, the trained linear regression (LIN) models can properly give out the potential energy evolution along the [100] direction with deviation less than 5%. By the dataset of given distances (9.3 Ĺ, 9.65 Ĺ and 10 Ĺ) between two Re monolayers in Gr/Re systems, LIN and Bayesian ridge regression (BR) models can quantitatively predict the potential energy evolution of unknown distances (9.2 Ĺ, 9.4 Ĺ, 9.5 Ĺ and 9.6 Ĺ). The predicted magnitudes of potential energy corrugations by BR model divert less than 3 meV Ĺ−2 from DFT calculations. The prediction results for extrapolated distances (9.0 Ĺ and 9.1 Ĺ) deviate notably, but the extension of training dataset effectively improves the predictive ability of ML models, especially for the LIN model. Thus, the supposed strategy could become an effective method to investigate the frictional characteristics of complex interfaces. [ABSTRACT FROM AUTHOR]

Metal-organic frameworks (MOFs) are a new class of nanoporous materials that are widely used in various emerging fields due to their large specific surface area, high porosity and tunable pore size. Its excellent chemical tunability provides a wide material space, in which tens of thousands of MOFs have been synthesized. However, it is impossible to explore such a vast chemical space through trial-and-error methods, making it difficult to achieve custom design of high-performance MOFs for specific applications. Machine learning (ML) is a powerful tool for guiding materials design and preparation by mining the hidden knowledge in data, and can even make prediction of material properties in seconds. This review aims to provide readers with a new perspective on how ML has been changing the research and development paradigm of MOFs. The four main data sources for MOFs and how to select the suitable features (descriptors) are firstly presented to enable the reader to quickly acquire data and carry out machine learning. Moreover, the application of ML in the development of MOFs is highlighted from the perspectives of performance prediction, rational design and intelligent synthesis. Finally, the future challenges and opportunities of combining ML with MOFs from the points of view of data and algorithms are proposed. This review will provide instructive guidance for ML-assisted MOFs research. [Display omitted] • A comprehensive literature review on a new paradigm of MOFs research is reported. • The current main data sources and descriptors of MOFs are classified and discussed. • Further challenges are proposed to provide direction for the development of machine learning in MOFs. [ABSTRACT FROM AUTHOR]

The study of piezocatalysis has become an important topic in piezoelectric research, especially for addressing environmental issues. However, the reliance on nanofabrication seriously hinders investigating materials with complex compositional design for higher performance and large-scale application. In this work, we use a machine learning strategy to efficiently sample the vast compositional space for ceramic powders with excellent piezoelectric response that we expect to impact piezocatalytic performance. The ceramics, synthesized by solid state reaction methods, belong to the multi-component system (Ba 1−x−y Ca x Sr y)(Ti 1−u−v−w Zr u Sn v Hf w)O 3 with target property d 33, the piezoelectric coefficient. The highest d 33 tends to occur within the phase boundary region with coexisting rhombohedral, orthorhombic and tetragonal phases, especially on the rhombohedral phase side. We select (Ba 0.95 Ca 0.05)(Ti 0.9 Sn 0.1)O 3 as it combines a relatively large d 33 with the fewest number of elements. Its sintered ceramic exhibits a high d 33 of 605 ± 14 pC/N, consistent with the machine learning prediction 633 ± 70 pC/N. The mechanically-ground ceramic powders have an excellent piezocatalytic activity with a degradation rate of (2.16 ± 0.28) × 10−2 min−1 for RhB dye solution, comparable to the performance of previously reported nanoparticles. Our work provides further insight into the nature of piezoelectricity in BaTiO 3 based ceramics, and affords an effective strategy for searching for superior piezocatalysts suitable for large-scale applications. We use machine learning to accelerate the discovery of ceramic particles prepared by conventional solid-state reaction method with high piezocatalytic performance in the multi-component ferroelectric system (Ba 1−x−y Ca x Sr y)(Ti 1−u−v−w Zr u Sn v Hf w)O 3. After establishing a regression model as a surrogate model, we combine the predicted values and the associated uncertainty of the model to guide the search for compositions with high d 33 in the vast search space. When the d 33 predictions are mapped in the phase diagram, the meaning of the composition-property relationship becomes very clear: a relatively large d 33 within the MPB region with the rhombohedral phase to one side. Then we select the system (Ba 0.95 Ca 0.05)(Ti 0.9 Sn 0.1)O 3 as it combines a large d 33 with the fewest number of elements. Its sintered bulk ceramic exhibits a relatively high d 33 of 620 pC/N and a piezocatalytic activity with a degradation rate of 24.5 × 10−3 min−1 for the RhB dye solution. [Display omitted] • Machine learning accelerated composition design towards superior piezocatalytic performance in BaTiO 3 -based ceramic powders. • Data mining reveals the highest d 33 occurs within the phase boundary region, especially on the rhombohedral phase side. • (Ba 0.95 Ca 0.05)(Ti 0.9 Sn 0.1)O 3 ceramic exhibits a highest degradation rate of 24.5 × 10−3 min−1 for RhB dye. [ABSTRACT FROM AUTHOR]

High temperature (HT) strength and room temperature (RT) ductility trade-offs are always unavoidable in the design of refractory high entropy alloys (RHEAs). Exploring the vast chemistry space to find optimally strong and ductile RHEAs remains a highly challenging task. Herein, we formulate a machine learning-based design strategy fusing uncertainty estimation and clustering analysis to explore a special alloy system (NbTaZrHfMo) for collaborative optimization of HT strength and RT ductility. Four non-equimolar alloys with a superior combination of HT strength, RT ductility and HT specific yield strength were discovered and synthetized. The influence of elements on mechanical properties are analyzed and an optimal composition range is identified based on model prediction. This work provides a general design approach enabling the concurrent optimization of conflicting properties with a small data-trained machine learning model, thereby producing a recipe to accelerate the discovery of desired RHEAs and other materials within a vast search space. In the present work, we formulate a data-driven strategy combining ML model prediction of alloy properties, EI-based uncertainty utility and clustering-based selection of alloy composition to systematically explore the NbTaZrHfMo chemical space to accelerate the design of RHEAs with a synergistic improvement of high temperature (HT) strength and room temperature (RT) ductility. We show how to accelerate the discovery of desired RHEAs in a vast chemistry space using small data-trained machine learning (ML) model. This approach is demonstrated as efficient, and four non-equimolar alloys with a superior combination of HT strength, RT ductility and HT specific yield strength are designed and synthetized, which are expected to serve as potential alternative materials for HT structural applications. This work provides a general design approach enabling the concurrent optimization of conflicting properties with a small data-trained machine learning model, thereby producing a recipe to accelerate the discovery of desired RHEAs and other materials systems within a vast search space. [Display omitted] [ABSTRACT FROM AUTHOR]

The infiltration of hydrogen can weaken the strength and toughness of materials, especially in high-strength steel. Considering the inherent trade-off between strength and toughness in steel design, balancing these two factors with the material's hydrogen embrittlement sensitivity has become a huge challenge in steel design. The long cycle nature of hydrogen embrittlement testing makes the iteration of traditional trial and error methods infeasible for steel development. This work adopts active learning algorithms based on Bayesian and Pareto optimization to iteratively optimize the chemical composition and processing technology of high-strength steel, in order to simultaneously improve its hydrogen filled tensile strength (UTS_H) and hydrogen filled total elongation (TEL_H). After two rounds of iterative optimization, the iterative optimization algorithm significantly improved the strength prediction of materials after hydrogen charging. The machine learning model found 16 types of steel from over 15 million unknown materials. Eight materials showed better comprehensive performance than known data, with one material having a tensile strength exceeding 1800 MPa and an elongation of 12.6%. Subsequent experiments showed that the optimization algorithm demonstrated an understanding of the impact of austenite content on hydrogen embrittlement sensitivity in multiple experiments, and designed high-strength steel resistant to hydrogen embrittlement through the austenite content of martensitic steel. • ML model accurately screens high-strength steels for hydrogen embrittlement resistance. • Iterative use of Bayesian modeling to minimize trial cycle time. • 8 excellent hydrogen embrittlement-resistant materials are screened from 1500 candidate spaces through 2 rounds of iteration. [ABSTRACT FROM AUTHOR]

Materials informatics employs machine learning (ML) models to map the relationship between a targeted property and various materials descriptors, providing new avenues to accelerate the discovery of new materials. However, the possible ML models and materials descriptors are numerous, and a rational recipe to rapidly choose the best combination of the two is needed. In the present study, we propose a systematic framework that utilizes a genetic algorithm (GA) to efficiently select the ML model and materials descriptors from a huge number of alternatives and demonstrated its efficiency on two phase formation problems in high entropy alloys (HEAs). The optimized classification model allows an accuracy for identifying solid-solution and non-solid-solution HEAs to be up to 88.7% and further for recognizing body-centered-cubic (BCC), face-centered-cubic (FCC), and dual-phase HEAs to reach 91.3%. Furthermore, by employing an active learning approach, several HEAs with largest classification uncertainties were selected, experimentally synthesized and phase-identified, and augmented to the initial dataset to iteratively improve the ML model. The method serves as a general algorithm to select materials descriptors and ML models for various material problems including classification and optimization of targeted properties. [ABSTRACT FROM AUTHOR]

We formulate a materials design strategy combining a machine learning (ML) surrogate model with experimental design algorithms to search for high entropy alloys (HEAs) with large hardness in a model Al-Co-Cr-Cu-Fe-Ni system. We fabricated several alloys with hardness 10% higher than the best value in the original training dataset via only seven experiments. We find that a strategy using both the compositions and descriptors based on a knowledge of the properties of HEAs, outperforms that merely based on the compositions alone. This strategy offers a recipe to rapidly optimize multi-component systems, such as bulk metallic glasses and superalloys, towards desired properties. In the present study, we proposed a data-driven approach combining machine learning, experimental design and feedback from experiment to accelerate the search for multi-component alloys with target properties. We demonstrated the efficiency of our approach by rapidly obtaining several alloys with hardness 10% higher than the best value in the original training dataset via only seven experiments. In Iteration Loop I, a machine learning surrogate model is trained to learn the property-composition, relationship, y i = f (c i) , with associated uncertainties. The model is applied to the vast unexplored space and a utility function is employed to recommend the most informative candidate for the next experiment, which balances the exploitation and exploration. Feedback from experimental synthesis and characterization allows the subsequent iterative improvement of the surrogate model. Iteration Loop II is essentially same as Iteration Loop I, except that a features pool was introduced to the Iteration Loop I and a surrogate model is trained from composition (c i) and the preselected physical features (p i), y i = f (c i , p i). We found that the approach using both the composition and the descriptors based on domain knowledge can more effectively accelerate material optimization compared to the approach using only the compositions. Image 1 [ABSTRACT FROM AUTHOR]

[Display omitted] • The interlayer sliding energy barriers of MoS 2 with polarization was predicted by the ML method. • A charge density normalization method was proposed to accelerate and simplify the preparation of training DFT charge density dataset. • The LR model can accurately fit the energy changes during sliding in both parallel and anti-parallel stacking modes of MoS 2. • The prediction error of sliding potential energy barrier can be well controlled within 5%. Machine learning is an effective method to predict the potential energy surface of interlayer sliding between two-dimensional (2D) materials but with less density functional theory (DFT) calculation burden. In this study, the machine learning method has been used to predict the sliding energy barriers between two MoS 2 layers with both parallel and anti-parallel stacking modes, respectively. The sliding energy barrier prediction of anti-parallel stacking modes with only structure factors as features deviates from the DFT calculations remarkably. However, the introduction of interlayer charge density as the prediction feature significantly improves the prediction accuracy of anti-parallel stacking modes. This study enlightens the precise prediction of sliding energy barriers of polarized 2D materials with machine learning method. [ABSTRACT FROM AUTHOR]

A phase diagram is a critical tool in materials science, but establishing it often requires a large number of experiments, especially for multi-component systems. In this work, we propose a machine learning strategy to accurately predict the phase diagram for the multi-component ferroelectric system (Ba 1−x−y Ca x Sr y)(Ti 1−u−v−w Zr u Sn v Hf w)O 3 by combining classification and regression methods. Based on literature data, we construct by classification a diagram that maps composition and temperature to the phase, and identify octahedral factor, Matyonov-Batsanov electronegativity, the ratio of valence electron number to nuclear charge and core electron distance (Schubert) for the A site and B site cations as the dominant physical descriptors. A neural network (NN) regression model is adopted to accurately predict phase transition temperatures, i.e. the phase boundaries, so that a phase diagram can be established in composition space. In the region of phase boundaries, the relative proportions of coexisting phases are also estimated by their prediction probability. The predicted phase labels, location boundaries and coexisting phase proportions are experimentally validated for the ceramic (Ba 0.96 Ca 0.02 Sr 0.02)(Ti 0.98−x Zr 0.02 Hf x)O 3. Our work provides an effective approach to establish phase diagrams for multi-component systems and also predict fine boundary information. [Display omitted] [ABSTRACT FROM AUTHOR]

In a previous paper (Dong et al., 2021), we showed that the equilibrium molecular dynamics (EMD) method can be used to compute the apparent thermal conductivity of finite systems. It has been shown that the apparent thermal conductivity from EMD for a system with domain length 2 L is equal to that from nonequilibrium molecular dynamics (NEMD) for a system with domain length L. Taking monolayer silicene with an accurate machine learning potential as an example, here we show that the thermal conductivity values from EMD and NEMD agree for the same domain length if the NEMD is applied with periodic boundary conditions in the transport direction. Our results thus establish an exact equivalence between EMD and NEMD for thermal conductivity calculations. • Large-scale MD simulations are performed with an accurate machine learned potential. • Exactly equivalent thermal conductivity in EMD and NEMD has been established. • The exact equivalence is explained in terms of phonon-boundary scattering. • We suggest that EMD serves as a feasible alternative method to NEMD in finite systems. [ABSTRACT FROM AUTHOR]

The coarsening of γ' precipitates in superalloys involves multiple factors and impacts the performance of mechanical properties, such as strength and creep resistance. Classical kinetic theory describes the coarsening of γ' under limited conditions, as it shows deficiencies in γ' precipitation with excessive coarsening. Finding effective factors underlying the kinetic behavior significantly affecting precipitate coarsening remains largely unsolved. We address this by using machine learning models to identify materials descriptors describing γ' precipitate coarsening in Co-based superalloys. Using descriptors that include Young's modulus difference and valence electron number, we obtain an explicit relation assisted by symbolic regression that fits the experimental γ' size data far better than classical kinetic theory in describing γ' coarsening. We infer that the Young's modulus difference dominates γ' coarsening mechanism in different Co-based superalloys. Moreover, alloys containing additional elements with large Young's modulus should favor smaller γ' precipitates size, thus providing guidance for designing advanced multicomponent Co-based superalloys with high γ' coarsening resistance. [Display omitted] We utilize a computer vision framework incorporating the deep learning-based convolutional neural networks to extract accurate microstructural information from γ/γ' microstructure images for analysis in classical kinetic theories and machine learning (ML) models. Coupled with the classical kinetic theory analysis, it can well fit the earlier coarsening of regular γ' precipitates in some Co-based superalloys. However, the classical kinetic theory can not give a reasonable fitting and explanation on the fast coarsening of irregular γ' precipitate, which fully exposes its deficiency in dealing with practical problems. We use ML models and proposed descriptor selection strategy by fitting experimental γ' size data to describe γ' precipitate coarsening using descriptors such as Young's modulus difference (σ E) and valence electron number (V E N), which we identify from ML models. The relation we propose for γ' size prediction equation in terms of aging temperature (T), aging time (t), σ E and V E N is far superior in fitting to the data than classical kinetic theory, and Co-based superalloys with large σ E value tend to follow the γ' coarsening predicted by LSW theory. For strengthening materials, this model for γ' size, which inversely depends on σ E , predicts that the addition of elements with large Young's modulus (E), such as W, Mo, Re and Ru will favor smaller γ' precipitates in Co-based superalloys. This provides a means to design multicomponent Co-based superalloys with high γ' coarsening resistance. [ABSTRACT FROM AUTHOR]

[Display omitted] • We have proposed an augmentation strategy for empirical formula based on machine learning mothed. • A novel empirical formula of Ms as the function of compositions was fitted based on an augmented dataset combined experimental data with predicted data. • Compared with previous empirical formula, the accuracy and robustness of the novel empirical formula we proposed was significantly improved without additional experimental cost. • This strategy offers a recipe to build empirical formula based on a small experimental dataset. Tuning the martensite transformation temperature through composition design has become an important way to broaden the applicable temperature range of shape memory alloys (SMAs). The empirical formula based on traditional statistics is a key reference for composition design. Due to the lack of experimental data, a large deviation may exist among the prediction results from the empirical formulas obtained by different data sources. In present work, we proposed an augmentation strategy of empirical formula based on a machine learning method to build the relationship between martensite transformation start temperature (Ms) and compositions in Cu-Al-based SMA system. A series of ML models were established by physical and chemical features and a Gaussian radial basis kernel function support vector machine (SVR.rbf) model was screened out based on mathematical and domain knowledge criteria. An augmented empirical formula of Ms as the function of compositions was fitted based on the abundant augmented dataset combined experimental data with predicted data by the SVR.rbf model. Compared with previous empirical formula fitted by small experimental dataset, the accuracy and robustness of the augmented empirical formula was significantly improved without additional experimental cost. This strategy offers a recipe to build empirical formula based on a small experimental dataset. [ABSTRACT FROM AUTHOR]

Solid solution strengthening (SSS) influences the exceptional mechanical properties of single-phase high entropy alloys (HEAs). Thus, given the vast compositional space, identifying the underlying factors that control SSS to accelerate property-oriented design of HEAs is an outstanding challenge. In the present work, we demonstrate a relationship derived in terms of the electronegative difference of elements to characterize SSS for HEAs. We propose a model which shows superior performance in predicting solid-solution strength/hardness of HEAs compared to existing physics-based models. We discuss applications of our SSS model to HEA design and predict alloys with potentially high SSS in the four alloy systems AlCoCrFeNi, CoCrFeNiMn, HfNbTaTiZr and MoNbTaWV. Our findings are based on the use of machine learning (ML) methods involving feature construction and feature selection, which we employ to capture salient descriptors. We identify descriptors and propose a model for solid solution strengthening (SSS) in high entropy alloys (HEAs). After data collection and feature engineering using diverse machine learning models, we identify salient descriptors, including electronegativity difference and shear modulus, influencing SSS in HEAs. The descriptors serve to formulate an SSS model with a simpler form to predict the strength of single-phase HEAs with greater accuracy compared to existing physics based SSS models. The descriptors also offer a perspective on SSS in terms of electron-associated interactions amongst the elements of the HEA. Finally, we predict HEAs with high SSS in four typical alloy systems. To motivate experiments, specific composition ranges are predicted which have high strengthening and not previously studied. They span a broad space of elements and their combinations [Display omitted]. [ABSTRACT FROM AUTHOR]

With aid of good materials descriptors, machine learning algorithms are able to accelerate the design of new materials and uncover underlying mechanisms. In the present study, we adopt machine learning methods to discover the most important materials descriptors for properties of ferroelectric materials. A regression study, in typical BaTiO 3 -based and K 1/2 Na 1/2 NbO 3 -based lead-free ceramics and lead-contained PbMg 1/3 Nb 2/3 TiO 3 -PbTiO 3 ceramics, screens out three important materials descriptors determining ferroelectricity from 46 potential descriptors. The three descriptors of Matyonov-Batsanov electronegativity, ratio of valence electron number to nominal charge and core electron distance (Schubert) are confirmed to be dominant as well in classification of perovskite compounds into antiferroelectrics or not. The classification based on these descriptors exhibit an excellent accuracy of 96%, much higher than that of traditional criterion (89%) using tolerance factor and Pauling electronegativity. Furthermore, we propose a machine learning strategy based on our descriptors to predict the phase coexistence. The prediction probability after bootstrapping provides an effective approach to distinguish the phase boundaries and predict the phase ratio of coexisted phases. In all, we identified materials descriptors for ferroelectric materials, which is helpful to reveal the structure-property relationship of ferroelectric materials and guide the design of better ferroelectricity and piezoelectricity. In the present study, we aim at identifying materials descriptors for ferroelectricity by machine learning. In order to ensure the validity and universality of the selected descriptors, we adopt our strategy to both regression and classification problems in ferroelectric materials. The problems include the regression of remnant polarization Pr in lead-free BaTiO 3 based ceramics, K 1/2 Na 1/2 NbO 3 based ceramics and lead-based perovskites of PbMg 1/3 Nb 2/3 TiO 3 -PbTiO 3 , and the classification of antiferroelectrics or non-antiferroelectrics. For each problem, we assembled a particular dataset, followed by building machine learning models respectively. By screening out features from a large number of potential descriptors, we identify the most important common descriptors determining ferroelectricity for the four problems. These descriptors are general for those regression and classification problems, and the classification based on these descriptors exhibit an excellent accuracy of 96%, much higher than that of traditional criterion (89%) using tolerance factor and Pauling electronegativity, which demonstrate the efficiency of our approach for building a more explicit structure-activity relationship, and provides the guidance for the accelerated design of ferroelectric materials. [Display omitted] [ABSTRACT FROM AUTHOR]