Track G
Big Data, Artificial Intelligence amd Machine Learning Methods for Accelerated Materials Discovery and Advancement
ABSTRACTS
Session G-1 Advances in machine learning principles, algorithms, descriptors and databases, machine learning approaches, their interpretability and potential pitfalls
G-1:IL01 Enabling the 4th Paradigm for Accelerated Materials Innovation
K.A. PERSSON, University of California at Berkeley, Berkeley, CA, USA
Fueled by our abilities to compute materials properties and characteristics orders of magnitude faster than they can be measured and recent advancements in harnessing literature data, we are entering the era of the fourth paradigm of science: data-driven materials design. We will exemplify data-driven materials innovation through a few case studies, in particularly focusing on the remaining bottlenecks in the design-synthesis-characterization loop. Recent advances in the science of synthesis, using the thermodynamic data present in the Materials Project (www.materialsproject.org), reaction networks and machine learning will be highlighted. Recognizing that the fuel of any data-driven or machine learning effort is robust and systematic data, we will discuss the avenues for incentivizing data infrastructure and connectivity.
G-1:IL02 Construct Exchange-correlation Functional via Machine Learning and Delta-learning Method
GUANHUA CHEN, The University of Hong Kong, Shatin, Hong Kong
Density-functional theory has been widely used in quantum mechanical simulations. Despite of its success, the universal exchange-correlation (xc) functional has been elusive. About twenty years ago, neural networks has been introduced to construct the xc functional or potential. Due to the emergence of deep learning, this effort has gained the renewed momentum in recent years. In this perspective, I review the early efforts to approximate the xc functional or potential with neural networks. A key challenge was the transferability from the knowledge learnt from small molecules to larger systems. The transferability problem was recently resolved by adopting quasi-local density-based descriptors, which is rooted rigorously in the holographic electron density theorem. I discuss then the recent developments to employ the deep-learning techniques to learn the exact xc functional. It is important that the high-level ab initio molecular energy and as well as the corresponding electron density are targeted for the training. All these efforts can be encompassed under a general framework. In addition, the delta-learning method will be discussed.
G-1:IL03 New Approaches to Predicting and Understanding the Electrochemical Stability of Inorganic Materials
J. MONTOYA, Toyota Research Institute, Los Altos, CA, USA
Electrochemical stability is a critical property of the materials we must develop in order to advance technologies that can help mitigate climate change, but methods to predict and understand electrochemical stability are not prevalent in the current paradigm of machine learning for materials science. In this talk, I will survey methods to predict electrochemical stability and to identify causes of electrochemical instability developed at the Toyota Research Institute. These include (1) advanced techniques of electrochemical stability mapping using 3-dimensional pourbaix diagrams, enabling the identification of critical concentrations that induce electrochemical stability, (2) modeling of electrochemical time-series data that can identify causal relationships between variables describing electrochemical systems and (3) machine-learning based prediction of long-running electrocatalyst durability experiments. Finally, we outline a vision of how these tools can be shared and used by the community to engineer more durable materials that will ultimately help enable the development of more reliable technology for carbon-neutrality.
G-1:IL04 Materials Discovery using Simulations and Deep Learning
A. Merchant, S. Batzner, M. Aykol, E.D. Cubuk, Google DeepMind, Mountain View, CA, USA
Despite the recent advances in physical simulations and machine learning, the exploration of novel inorganic crystals remains constrained by the expensive trial-and-error approaches. Recent developments in deep learning have shown that models can attain emergent predictive capabilities with increasing data and computation, in fields such as language, vision, and biology. In this talk, we will present our recent results on how graph networks trained at scale can reach unprecedented levels of generalization, improving the efficiency of materials discovery by an order of magnitude. We will show how the scale and diversity unlock surprising modeling capabilities for downstream applications, including predictions of crystalline stability, ionic conduction, and structural properties of amorphous materials.
G-1:IL05 Vibrational Properties of Inorganic Materials from High-throughput Density-functional Perturbation Theory and Machine-learning
G.-M. RIGNANESE, Institute of Condensed Matter and Nanosciences (IMCN), UCLouvain, Louvain-la-Neuve, Belgium
The progress in first-principles codes and supercomputing capabilities have given birth to the so-called high-throughput (HT) ab initio approach, thus allowing for the identification of many new compounds for a variety of applications. A number of databases have thus become available online, providing access to properties of materials, mainly ground‑state though. Indeed, for more complex properties (e.g., linear responses), the HT approach is still problematic because of the required CPU time. To overcome this limitation, machine learning approaches have recently attracted much attention. In this talk, I will review recent progress in materials informatics focusing on the vibrational properties of inorganic materials which play of key role in various physical phenomena such as thermal conductivity, superconductivity, or ferroelectricity. I will first present our HT calculations of the full phonon dispersion based on density functional perturbation theory. I will briefly introduce the OPTIMADE API that was developed for searching the leading materials databases using the same queries. Finally, I will review the MODNet framework for predicting materials properties and which is particularly well suited for limited datasets through the selection of physically meaningful features.
G-1:L06 Equivariant Tensor Network Potentials
M. HODAPP, Materials Center Leoben, Leoben, Austria; A. Shapeev, Skoltech, Moscow, Russia
The computational cost of many state-of-the-art machine-learning interatomic (MLIPs) potentials increases exponentially with the number of atomic features. Low-rank tensor networks can overcome exponential growth in complexity, however, it is often not easy to encode the model symmetries into them. Here, we propose a formalism for rank-efficient equivariant tensor networks (ETNs) that remain invariant under actions of SO(3). Using ETNs, we develop a new class of MLIPs that demonstrate superior performance over existing MLIPs.
G-1:L07 Utility of Transfer Learning in Computational Materials Science
S.G. GOPALAKRISHNAN1, R. DEVI1, K.T. BUTLER2, 1Dept of Materials Eng., Indian Institute of Science, Bengaluru, India; 2Dept of Chemistry, University College London, London, UK
Use of deep learning (DL) models in materials science is often limited by the lack of ‘large’ datasets, i.e., > 100,000 datapoints, that are either computed or measured. For instance, there are no large datasets available for several key performance determining metrics in energy applications, such as diffusivities in battery electrodes and carrier recombination rates in photovoltaics. On the other hand, there are reasonably large datasets available on formation enthalpies, computed band structures, and crystal structures across wide chemistries. Thus, if key chemical, compositional, and structural trends can be captured in available large datasets and subsequently transferred, it will enable the use of DL models in smaller datasets as well. Hence, in my talk, I will explore the utility of current transfer learning (TL) approaches that are available for computational materials science. Specifically, TL involves training a DL model on a larger dataset and then retraining a portion of the model on a smaller dataset. I will quantify the accuracy, transferability, and swiftness of learning in TL models compared to DL models that have been trained from scratch. Finally, I will focus on TL models that are precise and capture underlying physics which can be used for materials screening.
G-1:L09 Multiscale Study of the Electronic Structure of Halide Perovskites Slabs
A. CHARKIN-GORBULIN, D. Beljonne, C. Quarti, University of Mons, Mons, Belgium; I. Poltavsky, A. TkatcheNko, University of Luxembourg, Luxembourg
Ab initio methods give unprecedented accuracy in reproducing and interpreting structural, electronic, and optical properties for molecules, extended materials, and interfaces. Nonetheless, the computational demands associated with ab initio calculations impose stringent constraints on both the length and time scales that can be explored, usually within the limit of a few tens of Ångstrom/pico-seconds. Recent advancements however are paving the way for the development of less expensive materials modelling approaches. The emergence of machine learning-based force fields (MLFF) in particular is revolutionizing the field, affording, for instance, for an efficient and accurate modeling of potential energy surfaces (PES) across diverse physical systems. This pushed the limit of highly accurate atomistic simulations up to hundreds of nanoseconds and thousands atoms. This work uses MLFF to systematically and effectively explore the PES of slab models of CsPbI3 slab at finite temperature, addressing the complexities arising from the presence of surfaces. Obtained structures are then combined with large-scale electronic structure calculations to obtain a consistent description of the electronic properties of these materials, inherently accounting for thermal effects.
G-1:L10 Performance and Trustworthiness of Different AI models for Predicting Mechanical Properties of Steel Sheets
G. MILLNER1, L. Romaner2, D. Scheiber1, M. Mücke1, 1Materials Center Leoben Forschungs GmbH, Leoben, Austria; 2Montanuniversität Leoben, Leoben, Austria
Reducing CO2 emissions in steel production requires increased scrap usage, which introduces unwanted impurity elements. The impact of these foreign elements on the mechanical properties is in many cases not entirely understood and predicting the impurity effects on mechanical properties of steel from processing solely with physical models is not feasible. In this talk we present data-driven approaches to predict tensile strength, yield stress and other parameters of cold-rolled steel strip produced by voestalpine Stahl GmbH. The data includes a full chemical analysis, as well as many parameters measured during all working steps of the process and the resulting mechanical properties. We present our data pipeline to create a well-formed dataset fit for further modelling tasks. The strength and limitations of different model types applied to the available data and features will be presented. Special emphasis lies on the trustworthiness of the predictions, which can be obtained by including an uncertainty estimation of the predicted values and/or determine the impact of each feature in the model. The combination of uncertainty estimation and feature importance analysis provides the user with the ability to understand and trust the predictions.
G-1:IL11 Symmetry Constraints in Machine Learning Models of Electronic and Atomic Interactions
B. KOZINSKY, Harvard University, USA
Discovery and understanding of next-generation materials requires a challenging combination of the high accuracy of first-principles calculations with the ability to reach large size and time scales. We pursue a multi-tier method development strategy in which machine learning (ML) algorithms are combined with exact physical symmetries and constraints to significantly accelerate computations of electronic structure and atomistic dynamics. First, density functional theory (DFT) is the cornerstone of modern computational materials science, but its current approximations fall short of the required accuracy and efficiency for predictive calculations of defect properties, band gaps, stability and electrochemical potentials of materials for energy storage and conversion. To advance the capability of DFT we introduce non-local charge density descriptors that satisfy exact scaling constraints and learn exchange functionals called CIDER [1]. These models are orders of magnitude faster in self-consistent calculations for solids than hybrid functionals but similar in accuracy. On a different level, we accelerate molecular dynamics (MD) simulations by using machine learning to capture the potential energy surfaces obtained from quantum calculations. We developed NequIP [2] and Allegro [3], the first deep equivariant neural network interatomic potential models, whose Euclidean symmetry-preserving layer architecture achieves state-of-the-art data efficiency and accuracy for simulating dynamics of molecules and materials. In parallel, we implement autonomous active learning of interactions in reactive systems, with the FLARE algorithm that constructs accurate and uncertainty-aware Bayesian force fields on-the-fly from a molecular dynamics simulation, using Gaussian process regression [4]. These MD simulations are used to explore long-time dynamics of phase transformations and heterogeneous reactions.
[1] K. Bystrom, B. Kozinsky, arXiv:2303.00682 (2023)
[2] S. Batzner et al, Nature Comm. 13 (1), 2453 (2022)
[3] A. Musaelian, S. Batzner et al, Nature Comm. 14, 579 (2023)
[4] J. Vandermause et al, Nature Comm. 13 (1), 5183 (2022)
Session G-2 Virtual materials design and evaluation
G-2:IL01 Navigating Materials Space at Warp Speed
A. WALSH, Department of Materials, Imperial College London, London, UK
Traditional materials modelling workflows, even in the form of high-throughput approaches, are limited to small numbers of compositions and structures. I will present progress in materials informatics solutions for navigating a larger crystal chemical space. This includes techniques for compositional screening based on elemental features [1,2] and mapping from chemical formulae to three-dimensional crystal structures. A particular focus will be given to hand-built chemical filters that can reduce the magnitude of the search space and filters that are learned from data in the form of deep learning models based on crystal graphs. The performance of data-driven and domain knowledge-inspired approaches will be compared. Outstanding challenges in the field including robust synthesisability metrics [3] and generative crystal models will also be discussed.
[1] D. W. Davies et al, Computational Screening of All Stoichiometric Inorganic Materials. Chem 1, 617 (2016); [2] A. Onwuli et al, Element Similarity in High-Dimensional Materials Representations, Digitial Discovery 2, 1558 (2023); [3] K. Tolborg et al, Free Energy Predictions for Crystal Stability and Synthesisability, Digitial Discovery 1, 586 (2022)
G-2:IL02 Tackling Ion Transport and Interfacial Evolutions in Solid-state Batteries Machine-learning and Cluster Expansion Strategies
Z. Deng1, A.A. Panchal2, 3, W. Xie1, G.S. Gautam4, P. Canepa1, 2, 3, 1Department of Materials Science and Engineering, National University of Singapore, Singapore; 2Department of Electrical and Computer Engineering, University of Houston, Houston, TX, USA; 3Texas Center for Superconductivity, University of Houston, Houston, TX, USA; 4Department of Materials Engineering, Indian Institute of Science, Bengaluru, Karnataka, India
Computational material science is crucial to establishing a firm link between complex phenomena occurring at the atomic scale and macroscopic observations of functional materials, such as energy materials for rechargeable batteries. Storing and distributing green energy is central to the modernization of our society. Rechargeable batteries, including lithium (Li)-ion batteries, contribute substantially to shifting away from oil and other petrochemicals. Commercial Li-ion batteries suffer from stability issues. All-solid-state batteries utilizing solid-electrolyte “membranes” separating the distinct chemistries of the electrode materials appear to be a safer alternative. Nevertheless, stabilizing solid-solid “buried” interfaces in all-solid-state batteries remains a poorly understood aspect. In my talk, I will showcase the power of simulations to inform the complex reaction mechanisms, which take place at these complex interfaces. My talk will address two main aspects: 1) The advancement of first-principles kinetic Monte Carlo to study transport in fast-ion conductors. 2) I will showcase how machine-learned potentials can bring insight into the metal-anode/sulfide electrolyte interfaces, such as Li-metal/Li6PS5Cl.
G-2:L03 A Physics-informed Deep Learning Framework for Closed-loop Material Discovery
m. sharma Priyadarshini, O. Romiluyi, G. Wang, Department of Chemical and Biomolecular Engineering at The Johns Hopkins University, Baltimore, MD, USA; K. Miskin, Department of Materials Science & Engineering at The Johns Hopkins University, Baltimore, MD, USA; P. Clancy, Department of Chemical and Biomolecular Engineering at The Johns Hopkins University, Baltimore, MD, USA
A major factor contributing to the inefficiency of material discovery tools is the large combinatorial space of materials which can be sparsely observed for a given application. Searches of this space are often biased by expert knowledge and clustered close to known material configurations. Moreover, exhaustive experimental characterization or first principles calculations are extremely expensive, leading to small available data sets. Therefore, there is a need to develop algorithms that can efficiently search this large space. In this talk, we will introduce PAL 2.0, a method that combines a physics-informed belief model with Bayesian optimization (BO). The key contributing factor of PAL 2.0 is the creation of a physics-based hypothesis using XGBoost and Neural Networks which provide a physics-based prior to the Gaussian process model used in BO. Our method is unique since it picks out the physical descriptors that are most representative of the material domain, making the search unbiased toward expert knowledge. We will demonstrate PAL 2.0's performance on three computational semiconductor datasets. We will also present how PAL 2.0 is being used in a closed-loop setup with experimentalists to discover high-temperature shape memory alloys for space actuation applications.
Funding acknowledgment:
MSP, OVR and PC acknowledge support from the U.S. Department of Energy (DOE), Basic Energy Sciences (BES), under award DE-SC0022305. KM thanks Johns Hopkins University for his support. The authors acknowledge the support afforded by access to the computing 376 facilities at the petascale Advanced Research Computing at 377 Hopkins (ARCH) facility (rockfish.jhu.edu), supported by the 378 National Science Foundation (NSF), Grant Number OAC 379 1920103, for providing the extensive computational resources 380 needed here. Partial funding for the infrastructure for ARCH was originally provided by the State of Maryland.
G-2:L04 Inverse Design of Metal-organic Frameworks for Direct Air Capture of CO2 via Deep Reinforcement Learning
HYUNSOO PARK, S. Majumdar, X. Zhang, J. Kim, B. Smit, Imperial College London, London, UK
The combination of several interesting characteristics makes metal-organic frameworks (MOFs) a highly sought-after class of nanomaterials for a broad range of applications like gas storage and separation, catalysis, drug delivery, and so on. However, the ever-expanding and nearly infinite chemical space of MOFs makes it extremely challenging to identify the most optimal materials for a given application. In this work, we present a novel approach using deep reinforcement learning for the inverse design of MOFs, our motivation being designing promising materials for the important environmental application of direct air capture of CO2 (DAC). We demonstrate that the reinforcement learning framework can successfully design MOFs with critical characteristics important for DAC. Our top-performing structures populate two separate subspaces of the MOF chemical space: the subspace with high CO2 heat of adsorption and the subspace with preferential adsorption of CO2 from humid air, with few structures having both characteristics. Our model can thus serve as an essential tool for the rational design and discovery of materials for different target properties and applications.
G-2:IL05 Self-driving Fluidic Labs: Accelerated Materials Discovery, Optimization, and Manufacturing
M. ABOLHASANI, Department of Chemical & Biomolecular Engineering, North Carolina State University, Raleigh, NC, USA
The current human-dependent paradigm of experimental research in chemical and materials sciences fails to identify technological solutions for worldwide challenges in a short timeframe. Recent advances in reaction miniaturization, automated experimentation, and data science provide an exciting opportunity to reshape the discovery, development, and manufacturing of new advanced functional materials related to energy transition and sustainability. In this talk, I will present a Self-Driving Fluidic Lab for accelerated discovery, optimization, and manufacturing of emerging advanced functional materials with multi-step chemistries, through the integration of flow chemistry, online characterization, and artificial intelligence (AI). I will discuss how modularization of different synthesis and processing stages in tandem with a constantly evolving AI-assisted modeling and decision-making under uncertainty can enable resource-efficient navigation through high-dimensional experimental design spaces. Example applications of SDFL for the autonomous synthesis of clean energy nanomaterials will be presented to illustrate the potential of autonomous labs in reducing materials discovery timeframe from +10 years to a few months.
G-2:IL06 Machine Learning Discovery of Materials
J. SCHMIDT, Materials Theory ETH Zurich, Zurich, Switzerland; P. Borlido, Department of Physics, University of Coimbra, Portugal; A. Romero, Department of Physics and Astronomy West Virginia University, USA; T. Cerqueira, Department of Physics, University of Coimbra, Portugal; S. Botti, RC-FEMS and Faculty of Physics, Ruhr University Bochum, Germany; M. Marques, RC-FEMS and Faculty of Mechanical Engineering Ruhr University Bochum, Germany
Graph neural networks for crystal structures typically rely on atomic species and positions as input. We develop crystal-graph attention networks (CGATs) that substitute precise bond distances with graph distance embeddings. This approach enables high-throughput investigations based on both compositions and crystal structure prototypes. By integrating a newly curated dataset of 4 million materials with CGATs, we have efficiently explored thousands of prototypes searching through billions of materials. As a result, the size of the convex hull has more than doubled to over 100,000 materials. To enhance the accuracy of density functional theory (DFT) calculations regarding thermodynamic stability, we offer SCAN and PBEsol databases for stable and meta-stable materials. We research in detail the dependence of the prediction error when transfer learning from PBE to these datasets. Moreover, transfer learning enables us to extend our high-throughput searches to two-dimensional materials.
[1] Schmidt et al., Machine-learning-assisted determination of the global zero-temperature phase diagram of materials, Advanced Materials (2023)
[2] Schmidt et al., Crystal graph attention networks for the prediction of stable materials, Sci. Adv. 7.49 (2021)
G-2:L07 Accelerated Alloy Discovery and Optimization through the Batch-wise Improvement in Reduced Design Space using a Holistic Optimization Technique (BIRDSHOT)
r. arroyave, Texas A&M University, College Station, TX, USA
The Refractory High Entropy Alloy (RHEA) space is vast, and it is impossible to explore using conventional approaches to materials discovery. In this talk, I will present the Batch-wise Improvement in Reduced Design Space using a Holistic Optimization Technique (BIRDSHOT) framework. BIRDSHOT incorporates the strengths of ICME and combinatorial methods while addressing all their drawbacks, as it: (i) employs novel machine learning (ML) and data-driven search algorithms to identify efficiently the feasible regions amenable to optimization; (ii) exploits correlations to fuse simulations and experiments to obtain efficient ML models for predicting PSPP relations; (iii) uses Bayesian Optimization (BO) to make globally optimal iterative decisions regarding which region in the RHEA space to explore/exploit, leveraging existing models and data; (iv) is capable of carrying out multiple optimal parallel queries to the design space. We show how we have been using BIRDSHOT to search for next-generation refractory alloys for turbine engine applications.
Session G-3 Integrating machine learning and simulation for materials design and manufacturing
G-3:IL01 Machine Learning Guided High-throughput Combinatorial Printing and Characterization towards Autonomous Materials Discovery and Manufacturing
YANLIANG ZHANG, University of Notre Dame, Notre Dame, IN, USA
The convergence of advancements in machine learning and high throughput materials synthesis and characterization provides unprecedented opportunities to accelerate materials discovery and manufacturing for energy conversion and storage and develop innovative clean and sustainable energy technologies. In this talk, I will present our recent research breakthroughs on developing and synergistically integrating machine learning, high throughput combinatorial printing and flash sintering, and high-resolution scanning probe microscopy to accelerate the discovery of sustainable energy materials. First, I will present our work on developing machine learning integrated high throughput combinatorial printing methods to enable rapid discovery of high-efficiency and low-cost thermoelectric materials of record high performances for waste heat recovery and solid-state cooling. Second I will introduce machine learning guided autonomous and intelligent manufacturing of flexible thermoelectric and electronic devices. Finally, I will talk about our work in applying our advanced materials and manufacturing methods for developing innovative, clean and sustainable energy and sensor systems.
G-3:IL02 Material Discovery and Simulation using Machine Learning Potentials
SEUNGWU HAN, Department of Materials Science and Engineering, Seoul National University, Seoul, South Korea
The machine learning potential (MLP) derives total energies using machine learning models trained on reference DFT datasets. The accuracy and usefulness of MLPs have been demonstrated through numerous examples of large-scale and long-term simulations across diverse application domains, such as catalysts, batteries, and semiconductors. The data-centric approach in MLPs represents a radical departure from traditional DFT calculations or classical potentials, leading to a paradigm shift in atomistic simulations and paving the way for new technological advancements. In this presentation, we will explore applications of MLPs, which are enabled by the accuracy and speed of MLPs. First, we develop crystal structure prediction of completely unknown compounds by employing neural network interatomic potentials as a high-fidelity surrogate model of DFT. Using this approach, we explore unknown ternary metal oxides. In the second part of the presentation, we will discuss on the application of machine learning potential to the process simulation, such as dry and wet etching. In particular, we will present a robust protocol to sample chemical reactions.
G-3:L03 Structure Complements: A New Materials Taxonomy for ML-guided Materials Discovery
J.M. RONDINELLI, K.D. MILLER, Northwestern University, Evanston, IL. USA
We present a new paradigm of materials taxonomy, dubbed “structure complements,” by generalizing anti-structures (or inverse structures). Materials properties depend strongly on crystal geometry but also on the distribution of charge. Thus, our algorithm for classifying materials by geometry and cation/anion decoration proves not only useful as a novel categorization scheme but also as a framework for targeted materials discovery. As a use case, we showcase a workflow which combines structure complement analysis, a transparent machine learning model, and high throughput density functional theory calculations to discover novel ferroelectric materials. We then examine the microscopic origins of ferroelectricity in these new quasi-2D materials and compare them to state-of-the-art compounds. The workflow is designed to be integrated into an autonomous, closed-loop materials discovery platform which integrates a unified materials database, machine learning, simulation, and high-throughput synthesis and characterization.
G-3:L04 Bayesian Optimization of Carbide Free Bainitic Steels
B. SCHUSCHA1, D. Scheiber1, D. Brandl1, M. Mücke1, L. Romaner2, 1Materials Center Leoben Forschung GmbH, Leoben, Austria; 2Dept of Materials Science, Montanuniversität Leoben, Leoben, Austria
In response to increasing demands for high-performance structural materials, there is a need to develop steels that possess superior combinations of strength and ductility. A promising alloying concept for achieving this are carbide-free bainitic steels. We choose the Fe-C-Si-Mn-Cr-Mo-Al-Mo system, which results in a problem space of 8 dimensions. To effectively query this problem space and finding new chemical compositions, while minimizing the number of need samples and therefore the cost, Bayesian optimization is used in an adapted form. A graphical model is employed as the core of the Bayesian optimization. It embodies a probabilistic process-microstructure-property relationship which allows the integration of physical knowledge. Leveraging physical model, a modular structure is created that can be adapted which evolving knowledge. The dataset combines low-fidelity literature data and high-fidelity characterization data, which includes microstructure details and allows improvement of modules in this structure. We want to present you in this talk this approach to Bayesian optimization. With it, we hope do not only find a better process-microstructure-property-relationship, but also completely novel carbide free bainitic steel compositions.
G-3:IL06 Materials and Molecular Modelling, Imaging, Informatics and Integration (M3I3)
SEUNGBUM HONG, KAIST, Daejeon, South Korea
M3I3 is an algorithm to perform reverse engineering on future materials. M3I3 provides a means to achieve this goal effectively by mimicking the “reverse engineering” strategy with a higher level of creativity. M3I3 reverse engineers future materials of interest with superior performance, reliability, and minimum cost and environmental impact. Reverse engineering starts by analyzing the structure and composition of cutting-edge materials or products. Once we determine the performance of our targeted future materials, we need to know the candidate structure and composition of the future materials. This knowledge can only be available if we know the structure-property relationship of all materials and molecules at all scales. High-quality multi-scale and multi-dimensional experimental data will be the key to the success of our approach.
G-3:IL07 Design Metastability in High-entropy Alloys by Tailoring Unstable Fault Energies
WEI CHEN, Dept of Materials Design and Innovation, University at Buffalo, State University of New York, Buffalo, NY, USA
Metastable alloys with transformation-/twinning-induced plasticity (TRIP/TWIP) can overcome the strength-ductility trade-off in structural materials. Originated from the development of traditional alloys, the intrinsic stacking fault energy (ISFE) has been applied to tailor TRIP/TWIP in high-entropy alloys (HEAs) but with limited quantitative success. Here, we demonstrate a strategy integrating machine learning with materials modeling for designing metastable HEAs and validate its effectiveness by discovering seven alloys with experimentally observed metastability for TRIP/TWIP. We propose unstable fault energies as the more effective design metric and attribute the deformation mechanism of metastable face-centered cubic alloys to unstable martensite fault energy (UMFE)/unstable twin fault energy (UTFE) rather than ISFE. Among the studied HEAs and steels, the traditional ISFE criterion fails in more than half of the cases, while the UMFE/UTFE criterion accurately predicts the deformation mechanisms in all cases. The UMFE/UTFE criterion provides an effective paradigm for developing metastable alloys with TRIP/TWIP for an enhanced strength-ductility synergy.
G-3:L08 The MCL-MAP: A Platform for Accelerated Materials Design Based on Active Learning
J. SPITALER, D. Scheiber, N. Bedoya, Materials Center Leoben Forschung GmbH, Leoben, Austria; H. Tran, H. Gursch, Know Center GmbH, Graz, Austria; L. Romaner, Montanuniversität Leoben, Leoben, Austria
The MCL-MAP is a materials acceleration platform (MAP) based at Materials Center Leoben (MCL), aiming at radically new possibilities for optimizing and discovering high-performance materials. This is achieved by integrating physical modeling and machine learning (ML) in a hybrid modeling approach, where process-structure-property relationships are optimized using an Active Learning Loop (ALL). We will present the details of the MCL-MAP in terms of hard- and software for the platform backbone, the database with FAIR curated data, the framework for running physical modeling and Bayesian optimization algorithms, and the integrated software services for tasks such as interactive analytics, data visualization, data exploration and execution of modeling pipelines. Moreover, we will present the integrated literature search tool that enables semi-automated metadata extraction and digitization of data that are then saved with the original work. In addition, we are explaining the modular architecture implemented on a container-based infrastructure, making it easy to extend and adapt it for future developments. Finally, we present first results for two use cases: Bainitic steels with optimized mechanical performance and perovskite dielectrics with exceptional energy storage capability.
G-3:L09 Metastable Transition Metal Dichalcogenides from Machine Learning Force Fields
ZHENZHU LI, A. Walsh, Department of Materials, Imperial College London, UK
Transition metal dichalcogenides such as MoSe2 and TaSe2 are renowned for their intriguing electronic and optical properties. Crystal engineering has recently enabled the reproducible synthesis of metastable layered polytypes beyond the ground-state 2H configurations. Here we trained a machine learning force field for WSe2, from first-principles, that can describe both the ground-state and metastable phases. We probe the local and average structures and properties of these phases using molecular dynamics simulations, including the associated temporal distortions. Combined with data from experimental synthesis and characterisation, we will comment on the nature of the 1T of the 1T’ phases.
G-3:IL11 Machine Learning-driven Optimization of 3D Printing Composite Structures and Processes
SEUNGHWA RYU, Mechanical Engineering, Korea Advanced Institute of Science and Technology, Daejeon, South Korea
This presentation explores the systematic use of artificial intelligence (AI) in enhancing 3D printed structures and processes through two case studies. In the first half, a data-oriented design approach is introduced for the refinement of bio-inspired composites using a minimal dataset. This strategy leads to an ideal compilation of composite design solutions that effectively balances strength, toughness, and lightness. In the subsequent segment, a computer simulation model is constructed for the Digital Light Processing (DLP) 3D printing technique, addressing the resulting residual stress. This model facilitates the prediction of undesirable deformation in 3D printed structures post-DLP printing. By merging process simulation with experimental data through a data-driven optimization approach, we demonstrate that the aimed structure can be achieved with minimal unintended deformation.
G-3:L12 Composition and Property Prediction of Polymer-derived Silicon Oxycarbides
KATHY LU, University of Alabama at Birmingham, Birmingham, AL, USA; Yi Je Cho, Sunchon National University, South Korea
Polymer-derived silicon oxycarbide (SiOC) materials enable the formation of homogeneous microstructures and high temperature stable properties. However, the relationships between the processing parameters and microstructures/properties have not been clearly understood. In this study, a materials informatics approach was employed to the SiOC materials to analyze and estimate the relationships. Datasets were constructed from results of previously reported literature about SiOC. The correlation analysis provided processing parameter ranking regarding the corresponding influences on the properties and microstructures. Such an understanding can be later utilized for desired material fabrication. Machine learning models with high accuracy were proposed using the ranked features obtained from the correlation analysis. In addition, important points on the data collection, correlation analysis, and machine learning as well as limitations of the current dataset were discussed. The proposed workflow for the SiOC materials can be extended to different types of polymer-derived ceramics by incorporating various features and targets involved in the processing variables, microstructures, and properties.
G-3:L13 Machine Learning Point Defect Reconstructions
I. MOSQUERA-LOIS1, S.R. Kavanagh1, 2, D.O. Scanlon3, A. Ganose4, A. Walsh1, 5, 1Thomas Young Centre & Dept of Materials, Imperial College London, London, UK; 2Thomas Young Centre & Dept of Chemistry, University College London, London, UK; 3School of Chemistry, University of Birmingham, Edgbaston, Birmingham, UK; 4Thomas Young Centre & Dept of Chemistry, Imperial College London, London, UK; 5Dept of Physics, Ewha Womans University, Seoul, South Korea
Point defects are a universal feature of crystals. The standard approach of modelling them is, however, prone to missing the ground state atomic configurations associated with energy-lowering reconstructions from the idealised crystallographic environment. Missed ground states compromise the accuracy of calculated properties. To address this issue, we have developed an approach to navigate the defect configurational landscape using targeted bond distortions. Although effective, this approach requires a high number of Density Functional Theory calculations, making it impractical for high-throughput studies. Here, we tackle this limitation by employing a machine learning force field to explore the potential energy surface of defects and identify promising candidate structures. By learning common defect motifs, the surrogate model successfully extrapolates defect reconstructions to unseen compositions. It identifies the correct ground state structure for 92% of tested cases (unseen defects in unseen compositions), thereby reducing the number of required calculations by 78%. Overall, it significantly reduces the computational cost associated with exploring the defect configurational landscape, enabling efficient high-throughput studies of defects in crystals.
Session G-4 High throughput materials characterization and testing
G-4:IL01 A-lab: An Autonomous Laboratory for the Accelerated Synthesis of Novel Inorganic Materials
G. CEDER, University of California at Berkeley and Lawrence Berkeley National Laboratory, Berkeley, CA, USA
We will present the development and initial successes of A-lab: an AI-driven autonomous facility for the closed-loop synthesis of inorganic materials from powder precursors. All synthesis and characterization actions in A-lab, including powder mixing and grinding, firing, characterization by XRD and SEM, and all sample transfers between them are fully automated, leading to a lab that can synthesize and structurally characterize novel compounds within 10-20 hrs of initiation. The A-lab leverages ab-initio computations through an API with the Materials Project, historical data sets that are text-mined from the literature, machine learning for optimization of synthesis routes and interpretation of characterization data, and active learning to plan and interpret the outcomes of experiments performed using robotics. Over 17 days of continuous operation, the A-Lab successfully developed synthesis routes for 41 novel compounds from a set of 58 targets that were identified using large-scale ab-initio phase stability data from the Materials Project. Synthesis recipes were proposed by natural language models trained on the literature and optimized using an active learning approach grounded in thermodynamics. Analysis of failed syntheses provide direct and actionable suggestions to improve.
G-4:IL02 Autonomous Combinatorial Experimentation for Atomic Layer Synthesis
ICHIRO TAKEUCHI, University of Maryland, College Park, MD, USA
As a branch of machine learning, active learning has attracted much attention recently, as it can effectively help navigate experimental sequences in materials research in real time. We are incorporating active learning in combinatorial exploration of functional materials in autonomous modes. Autonomous experimentation can be used to reduce the number of required experimental cycles by an order of magnitude or more. The array format with which samples of different compositions are laid out on combinatorial libraries is particularly conducive to active learning. We have recently demonstrated autonomous control of unit cell-level growth of functional thin films implemented in pulsed laser deposition. Dynamic analysis of reflection high-energy electron diffraction images is used to autonomously navigate multi-dimensional deposition parameter space in order to rapidly identify the optimum set of growth parameters for fabricating the targeted materials phase. I will also discuss other autonomous experimentation projects we are carrying out. This work is performed in collaboration with M. Lippmaa, H. Liang, and A. G. Kusne. This work is funded by NIST, ONR, and AFOSR.
G-4:L03 Data-driven Material Exploration of Multi-element Substituted Fluorides toward high Conductivity
TETSUYA YAMADA1, 2, Y. Taketomi3, F. Hayashi1, K. Teshima1, 2, 1Faculty of Engineering, Shinshu University, Nagano, Japan; 2Research Initiative for Supra-Materials, Shinshu university, Japan; 3Graduate School of Science and Technology, Shinshu University, Japan
Barium fluoride (BaF2) has a fluorite-type structure and is attracting high attention as a solid electrolyte material for fluoride ion batteries that can be used under high voltage. Previous research has reported that the elemental substitution is effective in improving the ionic conductivity of BaF2. However, single element substitution only explore much limited experimental space in elements, so that it should be difficult to discover innovative conductive compositions. By replacing various elements at the same time, the experimental space can be largely expanded and the possibility of finding the optimal composition increases. Herein, we aimed to efficiently search for multi-element substitution compositions with high conductivity by creating a material-conductivity correlated exploring system under data-driven approach using large number of fluoride material database. In the presentation, we will focused on preparation of discovery of solid solution compositions in the multi-element substituted fluorides and understanding of the relationship between substitutional elements and solid solution formation.
G-4:L04 Machine Learning Aids High Throughput Material Characterization
Q. ALI, A. Kovacs, J. Fischbacher, H. Oezelt, M. Gusenbauer, D. Boehm, H. Moustafa, T. Schrefl, Christian Doppler Laboratory for magnet design through physics informed machine learning, Department for Integrated Sensor Systems, University for Continuing Education Krems, Wiener Neustadt, Austria; M. Yano, N. Sakuma, A. Kinoshita, T. Shoji, Advanced Materials Engineering Division, Toyota Motor Corporation, Susono, Japan; Y. Hong, T. Devillers, N.M. Dempsey, Institute Néel, Université Grenoble Alpes, CNRS, Grenoble INP, Grenoble, France
Characterization of materials is essential to understand the underlying physical processes occurred inside the system during synthesis and processing, for example. With advent of combinatorial sputtering, compositionally graded materials may be fabricated and in this work we study rare-earth based thin films. This allows to study the effect of both composition and processing on structural (phase) and magnetic (coercivity) properties of materials. Due to large number of samples efficient characterization methods are paramount. We present a high-throughput materials characterization framework employing two machine learning (ML) models to characterize NdLaCeFeB based isotropic thin films. In the first ML model we applied partial least square (PLS) regression to predict coercivity using the composition and X-Ray diffraction (XRD) spectra. In second step we use a ML model to identify phases in each sample. This model is trained on simulated XRD patterns by considering the possible experimental artifacts. Overall, this analysis helps in elaborating, which phase in the system is favourable for increase in coercivity and vice versa. The financial support by the Austrian Federal Ministry of Labour and Economy, and the Christian Doppler Research Association is gratefully acknowledged.
G-4:L07 Passive Ultrasonic Beamforming for Fast and Efficient Imaging of Solids
F. LANZA DI SCALEA, C. Huang, A.Z. Hosseinzadeh, Experimental Mechanics & NDE Laboratory, Department of Structural Engineering, University of California San Diego, La Jolla, CA, USA
Ultrasonic imaging, implemented with methods such as the Synthetic Aperture Focus (SAF), is a well established technique to provide quantitative information on reflectors internal to a test medium in both medical and engineering applications. The traditional technique employs a number of transducers that both transmit and receive ultrasonic waves, that are then beamformed to create an image of the medium. Traditionally, in order to improve the image contrast, several transmissions are used, whether using the entire array of transducers (such as the Full Matrix Capture – FMC mode) or using a subset of transducers (sparse arrays) to increase imaging speed. There exists an opportunity to further increase the imaging speed while also simplify the imaging hardware if the transducers are used primarily as receivers only. This can be accomplished by using deconvolution (or normalized cross-power spectrum) operators that extract the virtual Impulse Response Function (IRF) between two transducer receivers, in the presence of only a few limited transmitters (or no transmitters whatsoever with ambient excitations). An averaged normalized cross-power spectrum operator provides the bet robustness against incoherent noise. If the number of transmitters is minimized, there is then an opportunity to achieve fast beamforming with minimum hardware complexity.
Session G-5 Big data, machine learning and artificial intelligence moving towards next generation smart manufacturing and sustainable development
G-5:IL01 Collaborative Intelligence for Accelerated Development of Clean Energy Technologies
SHIJING SUN, University of Washington, Seattle, WA, USA
Innovation in energy storage and conversion is essential for addressing global challenges such as climate change. Artificial intelligence (AI) have emerged as a powerful tool to accelerate materials and systems development. but there are still challenges in realizing the potential of computational designs in the laboratory. One question increasingly get asked on self-driving labs is that 'will robots replace scientists?' In this seminar, I will discuss, rather than replacing researchers, how emerging technologies can augment and amplify human expertise, leading to unprecedented breakthroughs in energy materials, device and systems. As an experimentalist and materials data scientist with experience in both academia and industry, I will present examples of data-driven approaches that can address atomic-to-device level challenges in materials science. From high-throughput battery device testing to closed-loop inorganic materials design and open-ended exploratory synthesis, I will focus on how to predict experimental outcomes, explain results with interpretable machine learning, and design new experiments that incorporate physical knowledge into an automated framework, thereby guiding the sustainable development of new materials.
G-5:IL03 Utilizing Latent Space for Material Research and Development and Toward Digital Transformation
TETSUYA SHOJI, Toyota Motor Corporation, Advanced R&D and Engineering Company, Advanced Data Science Management Div. WAVEBASE project, Susono, Shizuoka, Japan
Recent rapid development of AI and data analytics technology bring us tools for extract deeper insight of material measurement data. Especially, using dimensionality reduction technique, or unsupervised learning, gives us important hint to access features that buried in raw measurement data acquired in material development. For example, according to textbook, XRD data may contain various information, e.g. phase, lattice parameters, existence of amorphous, nano-structural features and grain size etc.. Even it is so, most of researchers ignores most of information that include in XRD and just use it as qualitative information. When we use dimensionality reduction technique to XRD data that obtained during R&D cycle, one can recognize rich information are included in XRD spectra. Case of XRD is just one of example, we can extract information not only from spectrum but from microscope images using simple dimensionality reduction in q-space, or wave number, expression. This technique is useful to access insight of experimental data space. We deployed Material DX platform “WAVEBASE” on Cloud for easier access to deeper insight of complicated material measurement data.
G-5:IL04 FAIR Data for Accelerated Materials Discovery: The NOMAD Project
C. DRAXL, Physics Department and IRIS Adlershof, Humboldt-Universität zu Berlin, Berlin, Germany
Data-centric approaches are already changing materials science by complementing our traditional research. In fact, the enormous amount of data produced every day by the community, represents an invaluable gold mine. To turn this raw material into knowledge and value, requires, however, significant effort to manage, share, and publish data. The web-based platform NOMAD (https://nomad-lab.eu) provides a solution to this. Originally established as a computational materials-science repository, it is now significantly expanding its activities to include various sample-synthesis routes, a variety of experimental characterization techniques, as well as prototypical use cases from different research fields. The NOMAD data infrastructure not only contains core data management and electronic laboratory notebook functionalities, it automatically extracts rich metadata from supported file formats, normalizes and converts data, and provides a faceted search with materials-science specific filters. In its AI Toolkit, NOMAD integrates data-analysis and machine-learning tools. That way, NOMAD provides a unified way to make materials science FAIR, i.e., to Find, Access, Interoperate with, and Reuse millions of FAIR data from different data sources and workflows on the way to novel discoveries.
G-5:IL05 Accelerating Development of Materials with Artificial Intelligence and Machine Learning
J. SAAL, M. MUSTO, Citrine Informatics, Bad Wiessee, Germany
The development of novel materials and manufacturing processes can be time consuming and expensive due to the costs of experiments and the complexity of hierarchical process-structure-property-performance (PSPP) relationships that are unique within materials classes (e.g.,polymers, metals, ceramics). Artificial intelligence (AI)-driven materials design is particularly useful on 1) optimization exercises that are high-dimensional, 2) problems where PSPP cause- and-effect are unknown, and 3) projects with resource constraints. While a researcher using traditional methods typically estimate which inputs are the most important to vary and develop experiments accordingly, AI can systematically create machine learning (ML) models using all input combinations and then act as a copilot for the researcher, suggesting a series of experiments to explore the design space in the most efficient manner. This talk will explore critical questions in the use of AI/ML in materials development in a smart manufacturing environment, including how to assess performance of these methods in design exercises and challenges towards designing an efficient, scalable data infrastructure.
G-5:IL06 A Field Polarized by AI: How to Navigate the Conclusions and Delusions?
J.C. AGAR, Department of Mechanical Engineering and Mechanics, Drexel University, Philadelphia, PA, USA
Science traditionally uses data for decision-making. Previously, data was manageable for human analysis. However, the emergence of advanced sensing technologies has led to a flood of large, fast-moving data from varied sources, overwhelming traditional analysis methods. Despite advancements in AI and models like ChatGPT, their limitations are evident. AI excels in interpolation but struggles with extrapolation, often producing unrealistic outcomes when beyond their training data's scope. The rush towards AI adoption has created a misleading perception of its potential in ferroics, overshadowing significant technical innovations and transformative insights. We delve into the convergence of massive data growth and artificial intelligence, underscoring their combined strengths and weaknesses in augmenting decision-making processes, especially in the realm of data-driven infrastructure. We explore the co-design of experimental systems, encompassing algorithms, comprehensive software solutions, and hardware, all aimed at actualizing scientific machine learning. We then tackle the challenges inherent in applying machine learning within the context of ferroelectrics, where concepts like order, symmetry, and periodicity form essential semantic relationships. Our discussion extends to the development of high-availability computational frameworks designed for deploying resilient, self-repairing services in science, thus facilitating materials science at the exascale. Additionally, we spotlight innovations in parsimonious neural networks, adept at learning geometric transformations in reciprocal space. We emphasize the role of stochastic averaging in enhancing noise robustness, surpassing traditional algorithmic approaches. We apply these techniques to a variety of common synthesis and characterization techniques in ferroelectrics. Finally, we examine the progress in AI co-design, wherein algorithms are optimized for programmable logic (e.g., FPGAs) for real-time <1 ms inference. This optimization enables rapid, intelligent analysis, decision-making, and control on ultra-low-cost, low-power devices at unparalleled speeds. We illustrate how this approach is instrumental in real-time data analysis, data reduction, and dose-controlled imaging across various scientific platforms, including electron microscopy.