Session F53: AI and Materials II

Machine Learning Driven Automated Scanning Probe Microscopy for Material Discovery: Applications in Ferroelectric and Optoelectronic Materials

Scanning probe microscopy (SPM) has become a mainstay of the field of materials science. However, until now, the search for interesting functionalities in SPM experiments has been guided by auxiliary information to identify objects of interest and the exploration of physical mechanisms depends on human-based decision making, i.e., operators determine parameters for subsequent experiments according to the previous results. Meanwhile, machine learning (ML) has been applied to explore the physical mechanisms encoded in microscopy data. The combination of ML and SPM offers the opportunity of developing ML-driven automated SPM for the discovery of materials' functionality and mechanism in an automated manner. In this talk, I will discuss our development of ML-driven SPMs for learning the functionality and mechanism of ferroelectric materials and optoelectronic materials in an automated manner. We implemented three ML models in our SPM including a deep residual learning framework holistically-nested edge detection (ResHed) model, a deep kernel learning (DKL), and a hypothesis learning model. First, the ResHed model converts the stream image data into the semantically segmented objects of interest, then SPM can perform spectroscopic measurement on thus discovered objects automatically, allowing a systematic investigation of the discovered objects. Second, the DKL actively learns the relationship between structural elements in images and properties encoded in spectra during experiments. Third, the hypothesis learning method identifies the best physical models that can describe the material behaviours in an automated manner during the experiment. We implemented these approaches in SPM here, however, these approaches be adapted to apply to a broad range of physical and chemical experiments.   

Accelerating the Search for High-Performance, Novel Materials with Active Learning - An Example: Thermal Insulators

Active-learning frameworks have the potential to greatly accelerate the search for new materials. By balancing exploitation and exploration, these approaches can efficiently search through materials space and find the regions that are most likely to contain promising candidate materials [1]. Here we present an active learning framework, that uses an ensemble of expressions found by the sure-independence screening and sparsifying operator (SISSO) approach [2,3], and we domnstrate it for the example of discovering new thermal insulators. We statistically process the predictions of independent SISSO models to automatically select the most promising material candidates and then calculate their thermal conductivity, κ_L, using the ab initio Green Kubo method [4]. Using this approach we are able to find multiple new thermal insulators and gain insights into what is driving down their κ_L.

[1] A. G. Kusne et al. Nat. Comm. 11, 5966 (2020)

[2] R. Ouyang et al. Phys. Rev. Mater. 2, 083802 (2018)

[3] T. A. R. Purcell et al. J. Open Source. Softw. 7, 3960 (2022)

[4] F. Knoop, T. A. R. Purcell, M. Scheffler, and C. Carbogno submitted   

In-silico discovery of HER/OER multi-metallic Alloy electrocatalysts through Density Functional theory calculations and active learning and machine learning

Multi-metallic alloy catalysts have gained attention as means of finding better alternatives to the currently known pure and bimetallic catalysts. As multi-metallic alloys possess a diverse set of surface sites, tuning the distribution of active adsorption and reaction site is key in changing the catalytic performance. However, the combinatorial search of the multi-elemental space is expensive, with the different elemental selection and ratio of mixing being the cause.

Herein, we propose different machine learning frameworks for the efficient search of promising multi-metallic alloys in hydrogen evolution reaction (HER) and oxygen evolution reactions (OER), the two reactions needed to produce green hydrogen. In the first part, Active Learning algorithm is implemented, showing that when adequately searched, the adsorption energy, and consequently the activity of the multi-metallic catalyst can be predicted to within MAE accuracy of 0.1 eV, with only a small fraction of possible DFT calculation candidates. In the second part, the expansion of search space in alloy species is pursued through the utilization of the electronic density of states (DOS) information of catalyst atoms, in conjunction with graph neural networks for the structural information of the alloy surfaces. We show generalizability of the machine can be achieved through multi-task learning approach of DOS and adsorption energy prediction tasks.

    

Efficient Discovery of Air Separation Adsorbents via Multi-Fidelity Bayesian Optimization

Metal-organic frameworks provide seemingly endless choices of metal and organic species to tune small molecule adsorption behavior. Specifically, we examine MFU-4l and its variants as a platform for O2/N2 gas separation, an application where solid adsorbents promise to dramatically reduce the energy intensity of a traditionally costly and inefficient process. We enumerate a set of possible structure modifications of MFU-4l as combinations of metal species and organic ligands to develop a list of over 10,000 structures, and we screen their binding affinity for O2 and N2 using ab initio density functional theory calculations. To reduce computational effort, we use a Bayesian optimization approach with a multi-fidelity surrogate model that combines both ab initio and experimentally-obtained binding energy data. Through this work, we present candidate materials that can selectively separate O2 from N2 at ambient conditions with dramatically decreased energy cost relative to current adsorbents.   

Unveil the unseen: exploit information hidden in noise

Noise and uncertainty are usually the enemy of machine learning, noise in the training data leads to uncertainty and inaccuracy in the predictions. However, Wilson's Renormalization Group Theory tells us that noise within a physical system can determine its macroscopic state. This idea inspired us to develop a machine learning architecture that extracts crucial information out of the noise. The phenomenology computes and then utilizes uncertainty in one target variable to predict a second target variable. We first apply this formalism to crystalline PbZr_0.7Sn_0.3O₃, using the uncertainty in dielectric constant to extrapolate heat capacity, correctly predicting a phase transition that would otherwise be missed. We then apply the methodology to specify a concrete mix that has high resistance to carbonation, and another concrete mix that has low environmental impact. Both mixes also fulfil targets on the strength, density, and cost. The specified mixes are experimentally validated against their predictions. Our generic formalism enables the exploitation of uncertainty in machine learning, which has a broad range of applications in the physical sciences and beyond.   

Prediction of Crystal Symmetry Groups for Binary and Ternary Materials from Chemical Compositions using Machine Learning

Data-driven modeling becomes a fundamental and integral approach to conduct scientific work besides experiment, theory, and computation. It is applied in almost all scientific and technological fields. In material sciences, it is used to expedite the discovery of new materials with pre-specified properties. This shall allow the leap toward the next generation’s related sets of materials. In this work, starting only from the chemical formula, the elemental properties are utilized to develop an accurate predictive ML model for the crystallographic symmetry groups classification, including crystal systems, point groups, Bravais lattices and space groups [1,2]. The first step was generating a materials space of all possible ternary and binary compounds based on the common and uncommon oxidation states of 77 elements. The number of possible elemental combinations surpassed 600 million materials in total for both ternary and binary materials [1-3]. The average balanced accuracy of the predictive model exceeded 95% for all symmetry groups. The success of this work will contribute effectively to the advancement of materials science and discovery.

[1] Alsaui, Abdulmohsen, et al. "Highly accurate machine learning prediction of crystal point groups for ternary materials from chemical formula." Scientific reports 12.1 (2022): 1-10.

[2] Alsaui, Abdulmohsen A., et al. "Resampling techniques for materials informatics: limitations in crystal point groups classification." Journal of Chemical Information and Modeling 62.15 (2022): 3514-3523.

[3] Baloch, Ahmer AB, et al. "Extending Shannon's ionic radii database using machine learning." Physical Review Materials 5.4 (2021): 043804.

    

Statistical Modeling of Frictional Properties: a Machine Learning Approach

Controlling frictional forces is one of the major challenges in addressing the tribological response of interfaces. Besides, interpreting electronic features with microscopic tribological properties could provide a route to design lubricant additives. However, explicit evaluation of the frictional and adhesive properties of the interface is quite challenging since it requires expensive computational and experimental approaches. Consequently, statistical machine learning (ML) models have been developed to predict the adhesive, cohesive, and Van der Waals interaction energy. Using first principle methods, different stacking orders of structurally stable monolayers from the c2db database are simulated to generate the lowest energy structure of 760 bilayers. Utilizing the least absolute shrinkage and selection operator (LASSO), structural and interface elemental properties of these bilayers are identified as the potential descriptors for predicting the response variables. Different regression models have been developed using the selected descriptors, and their accuracies are compared to find the best-optimized model for predicting the frictional energies of bilayers. The high coefficient of determination (R2) and lower root mean square error (rmse) of the best ML model give the statistical interplay between the frictional force and elementary properties. Additionally, SHapley Additive exPlanations (SHAP) values have been used to interpret the selected descriptors and ML models.   

Machine Learning-Based Microstructure Prediction for Laser-Sintered Alumina

Predicting material’s microstructure is essential to link processing-structure-property (PSP) during advanced manufacturing. Usually, electron microscopy is used to characterize the microstructure at the needed scale. This process is time- and labor-consuming. In this work, we demonstrate a machine learning-based microstructure prediction that not only predicts the microstructure for unknown conditions but also predicts all features of microstructure (e.g., grain size, grain shape, porosity, etc.). We modified and improved the generative adversarial network (GAN) to be suitable for better microstructure predictions. Realistic SEM micrographs can be generated from a condition. When the condition is a processing parameter, we can correlate and predict the material’s microstructure under an unknown processing parameter. If the condition is a material’s property, such as hardness. We can predict material’s microstructure for an unknown hardness. We show that the GAN-based algorithm can predict more complicated features in the microstructure after it is modified with residual blocks. To evaluate the prediction accuracy, we use pre-trained convolutional neural network (CNN) to predict hardness from the GAN-generated SEM micrographs. We found CNN-predicted hardness matched well with the hardness that was used to predict the microstructure.    

Causal relations in determining functionalities in perovskite oxides

Machine learning (ML) methods to solidify understandings on material structures and functionalities have now become common in the physical sciences community. However, the in-built correlative nature of traditional ML techniques fails to capture the causal mechanisms driving any physical phenomena. Our study focuses on exploring fundamental atomistic mechanisms behind A-site cation ordering in double perovskite oxides. The origin of cation ordering has remained as a mystery for years since several factors such as cation radii and/or oxidation states, charge ordering, cooperative first order Jahn−Teller distortions of B cations (FOJT), A-site vacancies coupled with SOJT distortion, and tilt of BO₆/B′O₆ octahedra, contribute to it. Bringing in the causal intuitions with density functional theory calculations made it possible to not only pin down the necessary condition for tunable cation ordering but also establish quantifiable (previously unknown) stricture-property relationship between geometry, modes, and ordering. A discussion on polarization switching mechanism as understood from a combination of first-principles study and causal relations will be included in the presentation.   

Machine Learning Prediction of Perovskite Solar Cell Properties under High Pressure

Halide perovskites are promising solar cell materials due to their suitable bandgap range and high tunability. However, materials based on the organic-inorganic (MA)PbI₃ (MA = CH₃NH₃⁺) suffer a chemical instability issue to heat and moisture due to the volatile MA cation, while the all-inorganic Cs-based analogs present a phase instability challenge where the functional perovskite phases are unstable at ambient conditions and spontaneously convert into the thermodynamically stable non-perovskite phase. Therefore, stabilizing the perovskite phases at the room condition is crucial to achieving higher efficiency and commercialization. Tuning the structure by applying pressure and strain is an effective way to modify the stability and electrical properties of perovskite phases. In this work, we investigate the leading structural features that determine the material properties of the perovskites upon compression. We use various machine learning models to train the large-scale dataset obtained from first-principles DFT calculations. This study will provide insights into developing general models to predict the relationship between structural and electrical properties of similar perovskite structures using cost-effective machine learning approaches.   

Double descent, linear regression, and fundamental questions in materials model building

Though many data science concepts are just glosses on ideas that predate the data science revolution by years or even decades, some suggest altogether new approaches or raise fundamental questions. The phenomenon of double descent behavior in neural networks defies intuition and may seem to violate the "no free lunch" theorem. Is double descent behavior peculiar to neural networks? Or is it more general? We illustrate double descent in a simple linear regression model and then revisit basic questions in alloy model building, using the cluster expansion and machine learned interatomic potentials as illustrations. How is convergence impacted by the range of interaction? Or the order of an n-body interaction? How completely must we span configuration space with our expansions? We address these questions from the perspective of both mathematics and physics and discuss the implications for practical alloy models.   

Curie temperature prediction models of magnetic Heusler alloys using machine learning methods based on first-principles data from ab-initio KKR-GF calculations

Ordered and disordered magnetic Heusler alloys are an important class of materials in science and applications. Using Curie temperatures (T_c) of Heusler alloys calculated by the Korringa-Kohn-Rostoker Green function (KKR-GF) method and a subsequent Monte Carlo (MC) approach [1], we trained and evaluated several machine learning models to predict T_c based on atomic, magnetic, and structural properties. We studied multiple descriptor selection methods to determine the most meaningful physical quantities in the given phase space.

We compared the performance of regression and classification models in order to predict the range of the T_c of given compounds without performing the MC calculations. Since the MC calculation takes about as many computational resources as the ab-initio calculation, it would be favorable to replace either step with a less computational intensive method as e.g. machine learning. We discuss the necessity to generate the magnetic ab-initio results in order to make a quantitative prediction of the T_c.

This work can be seen as a small-scale case study in which lightweight machine learning algorithms can add value to existing ab-initio data and eventually replace costly computational steps in layered calculation workflows in the future.

[1] R.  Kovacik et al. (2022),  [10.24435/MATERIALSCLOUD:WW-PV]   

Physics Interpretable Ensemble Learning for Materials Property Prediction: Carbon as an Example

Machine learning models especially neural networks are most efficient and accurate methods to predict materials properties. But training a neural network model is time-consuming and often involves numerous parameters of little physical interpretability. Furthermore, input descriptors (if complicated) need to be designed manually to satisfy certain physical constraints such as of permutation, rotation, and reflection invariances. Here we propose an ensemble learning model consisting of regression trees to predict materials properties, using the formation energy and elastic constants of carbon allotropes as examples. Instead of using descriptors, our model adopts the computed energy and elastic constants from nine different classical interatomic potentials as inputs. The results of ensemble learning are more accurate than those from individual classical interatomic potentials, if density functional theory (DFT) results are used as the references. Because of the correlation between inputs and DFT reference, the regression trees can extract the relatively accurate formation energy or elastic constant that is calculated by the nine classical potentials and used as criteria for predicting the final properties. Our work shows the ensemble learning is applicable to different crystal structures while retaining physical interpretability to some extent.