Prediction Of Material Properties Research Articles

Interpretable machine learning‐assisted strategy for predicting the mechanical properties of hydroxyl‐terminated polyether binders

AbstractHydroxy‐terminated polyether (HTPE) binders are attractive in the weapons materials and equipment industry for their insensitive properties and flexibility. We propose an interpretable machine learning‐assisted modeling strategy to predict the mechanical properties of HTPE binders for the first time using machine learning methods. In this strategy, the effects of formulation composition, multiscale characterization, preparation conditions, and mechanical experimental conditions are evaluated on the mechanical properties of HTPE binders. As part of the study, three different techniques were used to predict material properties: bag‐based methods (Extra Random Tree, Random Forest), boosting‐based methods (XGBoost, CatBoost, and Gradient Boosted Regression), and Artificial Neural Networks (MLPs), all of which were highly accurate in predicting material properties. Based on this, SHAP analysis is used to explain how these influencing factors influence the material properties. An efficient method for examining HTPE binders formulations is provided by this strategy.

PyMSER─An Open-Source Library for Automatic Equilibration Detection in Molecular Simulations.

Automated molecular simulations are used extensively for predicting material properties. Typically, these simulations exhibit two regimes: a dynamic equilibration part, followed by a steady state. For extracting observable properties, the simulations must first reach a steady state so that thermodynamic averages can be taken. However, as equilibration depends on simulation conditions, predicting the optimal number of simulation steps a priori is impossible. Here, we demonstrate the application of the Marginal Standard Error Rule (MSER) for automatically identifying the optimal truncation point in Grand Canonical Monte Carlo (GCMC) simulations. This novel automatic procedure determines the point at which a steady state is reached, ensuring that figures of merit are extracted in an objective, accurate, and reproducible fashion. In the case of GCMC simulations of gas adsorption in metal-organic frameworks, we find that this methodology reduces the computational cost by up to 90%. As MSER statistics are independent of the simulation method that creates the data, this library is, in principle, applicable to any time series analysis in which equilibration truncation is required. The open-source Python implementation of our method, pyMSER, is publicly available for reuse and validation at https://github.com/IBM/pymser.

Universal Ensemble-Embedding Graph Neural Network for Direct Prediction of Optical Spectra from Crystal Structures.

Optical properties in solids, such as refractive index and absorption, hold vast applications ranging from solar panels to sensors, photodetectors, and transparent displays. However, first-principles computation of optical properties from crystal structures is a complex task due to the high convergence criteria and computational cost. Recent progress in machine learning shows promise in predicting material properties, yet predicting optical properties from crystal structures remains challenging due to the lack of efficient atomic embeddings. Here, Graph Neural Network for Optical spectra prediction (GNNOpt) is introduced, an equivariant graph-neural-network architecture featuring universal embedding with automatic optimization. This enables high-quality optical predictions with a dataset of only 944 materials. GNNOpt predicts all optical properties based on the Kramers-Krönig relations, including absorption coefficient, complex dielectric function, complex refractive index, and reflectance. The trained model is applied to screen photovoltaic materials based on spectroscopic limited maximum efficiency and search for quantum materials based on quantum weight. First-principles calculations validate the efficacy of the GNNOpt model, demonstrating excellent agreement in predicting the optical spectra of unseen materials. The discovery of new quantum materials with high predicted quantum weight, such as SiOs, which host exotic quasiparticles with multifold nontrivial topology, demonstrates the potential of GNNOpt in predicting optical properties across a broad range of materials andapplications.

From processing to properties: Enhancing machine learning models with microstructural information in polymer nanocomposites

For polymers and their composites, processing conditions and the resultant microstructures are crucial in determining their properties. Traditional machine learning (ML) methods typically focus on establishing direct relationships between processing parameters and material properties, often overlooking the critical intermediate step of how processing influences microstructure, limiting the predictive accuracy. In this study, we introduce an approach that first establishes a detailed relationship between processing parameters and the resultant microstructure, and then uses transfer learning and feature fusion to integrate this relationship into the prediction of material properties. Using carbon black-reinforced rubber composites (CRC) as an example, we compared ML models in predicting mechanical properties from processing data. A multi-task deep neural network performed best achieving an R2 of 0.763 with only processing data as input. When incorporating transfer learning and feature fusion, the R2 improved to 0.852 and 0.878, respectively. Shapley explanation analysis validated our approach, highlighting the importance of integrating processing, microstructure, and properties in ML models. This method emphasizes the critical effect of comprehensively considering processing–(micro)structure–property (P–S–P) relationships, leading to more accurate predictions in polymer composite research.

Bayesian optimization acquisition functions for accelerated search of cluster expansion convex hull of multi-component alloys

Atomistic simulations are crucial for predicting material properties and understanding phase stability, essential for materials selection and development. However, the high computational cost of density functional theory calculations challenges the design of materials with complex structures and composition. This study introduces new data acquisition strategies using Bayesian-Gaussian optimization that efficiently integrate the geometry of the convex hull to optimize the yield of batch experiments. We developed uncertainty-based acquisition functions to prioritize the computation tasks of configurations of multi-component alloys, enhancing our ability to identify the ground-state line. Our methods were validated across diverse materials systems including Co-Ni alloys, Zr-O compounds, Ni-Al-Cr ternary alloys, and a planar defect system in intermetallic (Ni1−x, Cox)3Al. Compared to traditional genetic algorithms, our strategies reduce training parameters and user interaction, cutting the number of experiments needed to accurately determine the ground-state line by over 30%. These approaches can be expanded to multi-component systems and integrated with cost functions to further optimize experimental designs.

Predicting the properties of bitumen using machine learning models trained with force field atom types and molecular dynamics simulations

This study enhances the molecular analysis of bitumen by transitioning from traditional chemical descriptors, such as SARA (Saturates, Aromatics, Resins, and Asphaltenes) fractions and elemental compositions, to specific force field atom types in Molecular Dynamics (MD) models. This shift improves the precision in predicting material properties critical for bituminous material characterization. Machine Learning Models (MLMs) were developed to use these atom types as input features, inherently reflecting fundamental chemical characteristics. Trained on data from over 1,770 LAMMPS simulations of diverse bitumen types and conditions, these MLMs enable the prediction of properties like density, heat capacity, solubility parameters, and thermal expansion coefficients without the need for additional MD simulations. The models utilize 30 chemical descriptors corresponding to specific atom types in the PCFF force field, which collectively account for over 95% of the influence on these properties. By accurately predicting fundamental, thermodynamic, and kinetic properties, the use of MLMs and force field atom types allows researchers to efficiently tweak the chemical nature of organic molecules and mixtures to achieve desired properties. With near-instantaneous prediction times, these MLMs offer valuable insights for advancing bitumen research in the construction and petroleum industries, reducing the need for more intensive simulation techniques.

Quantum support vector regression for predicting corrosion inhibition of drugs

This study evaluates the performance of Quantum Support Vector Regression (QSVR) in predicting material properties using limited data. Experimental results show that the QSVR model consistently produces superior prediction accuracy compared to previous conventional regression models. This improvement is especially evident in the prediction accuracy for small and complex datasets, where QSVR can better capture non-linear patterns. The superiority of QSVR in processing data with a quantum approach provides great potential in developing predictive models in materials science and computational chemistry.

High entropy alloys for hydrogen storage applications: A machine learning-based approach

Hydrogen is a clean energy carrier and has potential applications in energy storage, power generation, and transportation. This study explores the efficient and safe storage of hydrogen, particularly through solid-state methods using high entropy alloys (HEAs). HEAs have garnered attention for their versatility in tailoring properties for hydrogen storage. The integration of Machine Learning (ML) in designing HEAs offers an expedited approach, analyzing datasets and predicting material properties to enhance storage capacity, kinetics, and stability. Despite significant progress, the study acknowledges certain research limitations, particularly its relatively narrow focus on applying ML to HEAs for hydrogen storage. One of the biggest challenges with HEAs is their complexity, which.necessitates larger datasets to develop accurate predictive models. Collecting and analyzing existing HEA data for hydrogen storage using ML techniques is the main objective. Using algorithms like support vector regression (SVR), K-nearest (KNN), and random forest (RF), hydrogen-to-metal ratio (H/M) and valence electron configuration (VEC) are accurately predicted. The study proposes alloys with potential for HEA formation, identifying 741 with quaternary and 631 with quinary HEAs. These compositions are newly proposed and do not yet exist. Out of these, 774 HEAs are identified as potential candidates for hydrogen storage applications. Applying ML techniques, the selection process is more efficient, reducing the dependency on time-consuming experiments and making it easier to discover promising candidates.

Heteroscedastic Gaussian Process Regression for material structure–property relationship modeling

Uncertainty quantification is a critical aspect of machine learning models for material property predictions. Gaussian Process Regression (GPR) is a popular technique for capturing uncertainties, but most existing models assume homoscedastic aleatoric uncertainty (noise), which may not adequately represent the heteroscedastic behavior observed in real-world datasets. Heteroscedasticity arises from various factors, such as measurement errors and inherent variability in material properties. Ignoring heteroscedasticity can lead to lower model performance, biased uncertainty estimates, and inaccurate predictions. Existing Heteroscedastic Gaussian Process Regression (HGPR) models often employ complicated structures to capture input-dependent noise but may lack interpretability. In this paper, we propose an HGPR approach that combines GPR with polynomial regression-based noise modeling to capture and quantify uncertainties in material property predictions while providing interpretable noise models. We demonstrate the effectiveness of our approach on both synthetic and physics-based simulation datasets, including mechanical properties (effective stress) of porous materials. We also introduce an approximated expected log predictive density method for model selection, which eliminates the need for retraining the model during leave-one-out cross-validation, allowing for efficient hyperparameter tuning and model evaluation. By capturing heteroscedastic behavior, enhancing interpretability, and improving model selection, our approach contributes to the development of more robust and reliable machine learning models for material property predictions, enabling informed decision-making in material design and optimization.

Bayesian inverse inference of material properties from microstructure images

In this paper, we introduce a Bayesian framework designed for inverse inference, aiming to predict material properties/process parameters from microstructure images. The integration of Bayesian inference techniques with deep generative models establishes a robust tool for applications in materials science, particularly in material characterization and property control. This integration provides a novel approach to clarifying the reliability of predictions. The application of this framework to a sample problem involving the prediction of material properties from artificial dual-phase steel microstructures demonstrates its capability to estimate these properties while accounting for prediction uncertainties. Moreover, even in comparison to conventional regression methods in terms of point estimation, the proposed framework exhibits superior accuracy in prediction. These results clearly illustrate that the framework presented in this paper constitutes a powerful tool for achieving efficient material design.

Materials Properties Prediction (MAPP): Empowering the Prediction of Material Properties Solely Based on Chemical Formulas.

Predicting material properties has always been a challenging task in materials science. With the emergence of machine learning methodologies, new avenues have opened up. In this study, we build upon our recently developed graph neural network (GNN) approach to construct models that predict four distinct material properties. Our graph model represents materials as element graphs, with chemical formulas serving as the only input. This approach ensures permutation invariance, offering a robust solution to prior limitations. By employing bootstrap methods to train this individual GNN, we further enhance the reliability and accuracy of our predictions. With multi-task learning, we harness the power of extensive datasets to boost the performance of smaller ones. We introduce the inaugural version of the Materials Properties Prediction (MAPP) framework, empowering the prediction of material properties solely based on chemical formulas.

Rational Design of Deep Learning Networks Based on a Fusion Strategy for Improved Material Property Predictions.

The success of machine learning in predicting material properties is largely dependent on the design of the model. However, the current designs of deep learning models in materials science have the following prominent problems. First, the model design lacks a rational guidance strategy and heavily relies on a large amount of trial and error. Second, numerous deep learning models are utilized across various fields, each with its own advantages and disadvantages. Therefore, it is important to incorporate a fusion strategy to fully leverage them and further expand the design strategies of the models. To address these problems, we analyze that the main reason is the lack of a new feedback method rich in physical insights. In this study, we developed a feedback method called the Chemical Environment Clustering Vector (CECV) of compounds at different thresholds, which is rich in physical insights. Based on CECV, we rationally designed the Long Short-Term Memory and Gated Recurrent Unit fused with Deep Convolutional Neural Network (L-G-DCNN) to explore the field of structure-agnostic material property predictions. L-G-DCNN accurately captures the interactions between elements in compounds, enabling more accurate and efficient predictions of the material properties. Our results demonstrate that the performance of the L-G-DCNN surpasses the current state-of-the-art structure-agnostic models across 28 benchmark data sets, exhibiting superior sample efficiency and faster convergence speed. By employing different visualization methods, we demonstrate that the fusion strategy based on CECV significantly enhances the comprehension of the L-G-DCNN model design and provides a fresh perspective for researchers in the field of materials informatics.

Phase quantification using deep neural network processing of XRD patterns.

Mineral identification and quantification are key to the understanding and, hence, the capacity to predict material properties. The method of choice for mineral quantification is powder X-ray diffraction (XRD), generally using a Rietveld refinement approach. However, a successful Rietveld refinement requires preliminary identification of the phases that make up the sample. This is generally carried out manually, and this task becomes extremely long or virtually impossible in the case of very large datasets such as those from synchrotron X-ray diffraction computed tomography. To circumvent this issue, this article proposes a novel neural network (NN) method for automating phase identification and quantification. An XRD pattern calculation code was used to generate large datasets of synthetic data that are used to train the NN. This approach offers significant advantages, including the ability to construct databases with a substantial number of XRD patterns and the introduction of extensive variability into these patterns. To enhance the performance of the NN, a specifically designed loss function for proportion inference was employed during the training process, offering improved efficiency and stability compared with traditional functions. The NN, trained exclusively with synthetic data, proved its ability to identify and quantify mineral phases on synthetic and real XRD patterns. Trained NN errors were equal to 0.5% for phase quantification on the synthetic test set, and 6% on the experimental data, in a system containing four phases of contrasting crystal structures (calcite, gibbsite, dolomite and hematite). The proposed method is freely available on GitHub and allows for major advances since it can be applied to any dataset, regardless of the mineral phases present.

Effects of K, Rb, and Br doping on Cs3Bi2I9 perovskites: A design of experiments approach

This work presents the effect of doping with different species on the structural sites of lead-free halide perovskites, aiming to evaluate the potential of the produced materials as photoabsorbing layers in 3rd generation solar cells. Lead-free perovskite films of Cs3Sb2I9, CsCu2I3, Cs3Bi2I9, and CsPbI3 were synthesized via the spin-coating method and characterized using X-ray diffraction, scanning electron microscopy with chemical microanalysis (SEM-EDS), and Uv-Vis spectroscopy. Computational models based on Density Functional Theory (DFT) were used to simulate the structures and properties of some of the perovskites synthesized in this study, in order to validate the experimentally obtained data. The Cs3Bi2I6Br3 perovskite was subjected to modification through A-site doping with K and Rb and X-site doping with Br, with a comprehensive study on the effects of processing variables, anti-solvent quantity, and crystallization temperature using a 23 experimental design. The outcomes highlighted the critical role of the anti-solvent quantity in structure formation and electronic properties. Toluene, as an anti-solvent, produced unique morphologies not commonly reported. Computational studies were performed using Density Functional Theory (DFT) to complement the experimental findings. The optimization process revealed that Cs3Bi2I9 structures mainly constituted [CsI12] and [BrI6] clusters, with Cs–I and Bi–I average bond lengths consistent with experimental results. X-site Br doping induced slight distortion, while A-site Rb/K doping showed minimal lattice parameter changes. Notably, K-induced morphological alterations suggested potential electronic transformations. From an electronic structure perspective, Cs3Bi2I9 had a band gap of 2.07 eV, slightly increased to 2.10 eV with Br doping. Rb maintained Br-doped structure, while K significantly reduced the band gap to 1.96 eV. These computational results align well with our experimental data, showcasing the accuracy of our calculations in predicting material properties. The results confirmed that the perovskites selected through experimental design and validated by computational simulation are promising semiconductor materials for photovoltaic applications, as well as being a lead-free photoabsorbing material option.

(Invited) Parameter Prediction in Colloidal Quantum Dot Solar Cells Using Residual Neural Networks Trained on Experimental Data

Colloidal quantum dots (CQDs) are an attractive third-generation material for photovoltaics, especially as infrared materials for multi-junction solar cells. Characterization and materials parameter prediction both play a critical role in the rising efficiencies in this field. However, there are numerous materials properties that need to be measured in order to fully characterize a device, leading to high research costs and slow development times. Here, we leverage recent advancements in machine learning (ML), along with a recently-demonstrated multimodal spectral characterization method [1], to predict complex materials parameters in CQD solar cells from simple current density-voltage (JV) curves, using an algorithm trained on experimental data.[2]Past work ML prediction of materials properties has focused on using simulated data due to the difficulty in collecting enough experimental data to fully train an algorithm. While it is not feasible to fabricate thousands of devices in a typical research lab, it is possible to measure several thousand points on a single device. We used a micrometer-resolution 2D characterization method with millimeter-scale field of view to acquire enough training data using only a handful of devices, while learning the spatial relationships between materials parameters. Photoluminescence, trap state density, transient photovoltage, transient photocurrent, and carrier mobility were predicted (Figure 1) by five different residual neural networks, each using a ResNet block and containing over 300k learnable parameters. The design of the networks were based on the ResNet50 architecture [3]. We also implemented a novel neighborhood approach to reduce errors and account for spatial correlations in device behavior, leading to insights into transport mechanisms and correlation lengths in CQD thin films. This method could be used to study a wide range of material systems and devices, ultimately leading to a universal prediction model that greatly simplifies the characterization process for optoelectronic devices and accelerates development times.

Machine learning accelerated the prediction of mechanical properties of copper modified by TMDs based on molecular dynamics simulation

In this study, we constructed a dataset of elastic modulus and ultimate stress for copper material enhanced by Transition Metal Dichalcogenides (TMDs) through Molecular Dynamics (MD) simulations. Subsequently, leveraging chemical insights, we selected appropriate descriptors and established machine learning prediction models for elastic modulus and ultimate stress, respectively. Finally, the performance of the machine learning models was evaluated using a test set. The results demonstrate excellent performance of the machine learning models in predicting material properties. This work presents a novel approach for efficient material screening, demonstrating the synergy between MD simulations and machine learning in advancing materials research and intelligent material selection platforms.

Prediction of time-dependent concrete mechanical properties based on advanced deep learning models considering complex variables

Evaluating the mechanical properties such as compressive strength, tensile strength, and modulus of elasticity (MOE) of concrete is crucial for the design, construction, and quality control of engineering projects. Developing a comprehensive theoretical-based numerical model to predict the concrete mechanical properties is considered challenging, owing to the diversity of admixtures and curing environment affecting the performance of modern concrete. The deep learning (DL) model gives a feasible approach as data-driven methods for accurate prediction of various material properties these years. However, predicting time-dependent mechanical properties of modern concrete with complex mix constituents by DL algorithms is still confined so far. In this article, artificial neuron network (ANN) and long short-term memory (LSTM) based DL models were established to predict concrete compressive strength, tensile strength, and MOE at different ages, considering the mix design, curing condition, curing time and tension testing method as input variables. LSTM outperformed ANN in terms of the statistical indicators for accuracy evaluation with high training and testing time costs. Bidirectional modification and multi-head attention mechanism were implemented to further improve the prediction accuracy of LSTM. It is shown that the multi-head attention bidirectional LSTM had more precise predicting results (R2=0.9355 for compressive strength, R2=0.8930 for tensile strength, R2=0.9584 for MOE) and fitting of the mechanical property-age curves. The LSTM-based models provided a promising prediction tool for time-dependent concrete mechanical properties with complex input parameters.

Quantum-Chemistry Enabled AI for Materials Discovery for Energy Storage

A priori and reliable simulations can enable timely and cost-efficient design and discovery of materials for energy. Therefore, ‘Let’s Start from Computing’ is an optimal approach to initialize modern day R&D processes. In energy storage, beyond lithium-ion (BLI) research has the potential to revolutionize consumer electronics including portable and stationary power, transportation, and grid energy storage sectors. Multi-valent (Mg, Ca, Zn) energy storage or economically viable monovalent (Na, K) batteries, high-density metal-air, metal-sulfur batteries, or grid-storage systems are considered in the beyond lithium-ion research and development. All these R&D efforts require significant fundamental knowledge via a priori computations for materials discovery, property prediction, and optimization. Atomistic modeling when coupled with reliable Artificial Intelligence (AI) approaches can provide accurate insights to accelerate discovery of optimalelectrolytes, electrodes, and membranes for BLI systems to reduce the cost. Thus, coupled with AI and multi-scale simulations techniques, atomistic modeling can address prediction of molecular level properties of materials (redox potentials, solvation, spectroscopic, and reactivity) to down-select optimal materials or material combinations. In this presentation, I will describe some of our recent efforts in active learning coupled with large scale first principles simulations to down select/optimize desired molecules for flow battery technology. This concept can be utilized for design of experiments using autonomous experimentation. Additionally, I will describe some of our quantum chemistry-informed molecular property predictions redoxmers and liquid organic hydrogen carriers. In addition to molecules, I will present. A data driven approach to study longer time scale diffusion of ions for multivalent battery concepts. Figure 1

Introducing a new correlation functional in density functional theory

The correlation functional holds significance in density functional theory as it addresses electron–electron interactions beyond the mean-field approximation, enhancing the accuracy of total energy calculations, electronic excitations, and the prediction of materials properties. There are several expressions to describe this energy, and each of them has a unique set of errors in calculating particular properties of materials. This work offers a new correlation functional by employing the density's dependence on ionization energy. We theoretically derived this functional and combined it with the previously reported ionization energy dependent exchange functional to investigate its effect on the total energy, bond energy, dipole moment, and zero-point energy of 62 molecules. The comparison of this new functional in respect to existing widely used correlation models including QMC, PBE, B3LYP and Chachiyo models shows how well it works in producing accurate results with minimal mean absolute error.

Differential impedance analysis — Extensions and applications in machine learning

The technique of differential impedance analysis (DIA) has shown promising results in identifying appropriate model orders when applied to electrochemical impedance spectroscopy (EIS) measurements of a given medium. However, even with this method it remains challenging to reliably deduce general material properties of the medium from impedance data alone. Here, we discuss a number of possible extensions and modifications of the technique and, in particular, an extension of the process from mere model order identification to a complete modelling approach. In addition, the combination of DIA and machine learning methods to predict material properties is explored. Our results were validated with impedance measurements between 20Hz and 1MHz of moulding sand containing varying amounts of quartz and chromite sand as well as bentonite and carbon.