Bulletin of the American Physical Society
APS March Meeting 2023
Volume 68, Number 3
Las Vegas, Nevada (March 5-10)
Virtual (March 20-22); Time Zone: Pacific Time
Session N53: AI and Materials III |
Hide Abstracts |
Sponsoring Units: GDS DMP Chair: Trevor Rhone; Ayana Ghosh, Oak Ridge National Lab Room: Room 307 |
Wednesday, March 8, 2023 11:30AM - 11:42AM |
N53.00001: Combining generative modeling and genetic algorithm for atomistic structure search Venkata Surya Chaitanya Kolluru, Davis G Unruh, Joshua T Paul, Maria K Chan We developed a multi-objective genetic algorithm, FANTASTX (Fully Automated Nanoscale to Atomistic Structures from Theory and eXperiment) software, where we combined the theoretical tools with experimental data to determine the atomistic structure of experimentally observed materials. This work enhances the FANTASTX software with generative models to discover new candidate structures. We use variational autoencoders (VAE) and generative adversarial networks (GAN) to generate low-energy structures by training the models with system specific data. By combining the generative models with a genetic algorithm, we effectively sample local minima from regions with known data while exploring the potential energy landscape with genetic operations. Further, we train a crystal graph network to predict the formation free energies of the newly generated structures with the same training data. We implement different structure representation methods in the FANTASTX framework to test its effectiveness. We test the accuracy of the reconstruction of various examples from x-ray and electron microscopy data with different structure representation methods. |
Wednesday, March 8, 2023 11:42AM - 11:54AM |
N53.00002: Using Density Functional Theory and Machine Learning to Predict the Binding Energies of Metal-organics to Organic Functional Groups for Hybrid Material Creation Yifan Liu Understanding chemical interactions between organic and metal-organic molecules has wide-ranging interest to the vapor deposition community for creating hybrid organic-inorganic materials via techniques such as molecular layer deposition and vapor phase infiltration (VPI). In the case of VPI, a vapor-phase metal-organic precursor is infused into the bulk of a polymer and becomes incorporated at the nanoscale through either chemical interaction with the polymer or the formation of a non-volatile species via the introduction of a co-reactant. VPI has applicability in a number of industrially relevant fields including the creation of novel organic-inorganic hybrid membranes which have shown enhanced stability in organic solvents, while retaining high permeance and selectivity. Motivated by this application, this work uses density functional theory (DFT) to explore chemical interactions occurring during the VPI of polymer of intrinsic microporosity (PIM-1, a polymeric membrane material)with trimethylaluminum(TMA) and its co-reaction with water. These computations revealed that the coordination between the polymer and metal-organic is a critical mechanism for the formation of the hybrid and its resultant solvent stability.To expand understanding of this critical characteristic and accelerate the design of organic-inorganic hybrid materials, a DFT dataset of computed binding energies was generated from suitable and representative atomic-level models of several common polymer functional groups and over 100 metal-organic precursors. From this dataset, a predictive machine learning model for the binding energy of metal-organic molecules to polymers has been developed. This predictive model, along with the chemical guidelines obtained from feature analysis, will aid the selection of potential candidates for novel organic-inorganic hybrid membranes as well as hybrid material creation as a whole. |
Wednesday, March 8, 2023 11:54AM - 12:06PM |
N53.00003: Self-Supervised Learning for Material Property Prediction Rishikesh Magar, Amir Barati Farimani Machine learning (ML) models have successfully been used to predict material properties. However, the large labeled datasets required for training accurate ML models are elusive and computationally expensive. Recent advances in Self-Supervised Learning (SSL) frameworks capable of training ML models on unlabeled data mitigate this problem and demonstrate superior performance in computer vision and natural language processing tasks. Drawing inspiration from the developments in SSL, we introduce a new framework Crystal Twins. Our framework is a generic SSL methodology that can leverage unlabeled data for crystalline materials property prediction tasks. CT adopts a twin Graph Neural Network (GNN) and learns representations by forcing graph latent embeddings of augmented instances obtained from the same crystalline system to be similar. We implement Barlow Twins and SimSiam frameworks for self-supervised learning in CT. By sharing the pre-trained weights when fine-tuning the GNN for downstream tasks, we improve the performance of GNN on 14 challenging material property prediction benchmarks. |
Wednesday, March 8, 2023 12:06PM - 12:18PM |
N53.00004: Development strategies and hyperparameter optimization of Deep Learning potentials for multi-component and multi-phase Nickel-based Superalloys Marium Mostafiz Mou, Matthew J. Kindhart, Jared L Shortt, Ridwan Sakidja In this study, we developed Deep Learning interatomic potentials to model a multi-phase and multi-components system of Ni-based Superalloys. The complex system has up to ten elements with three major phase constituents, namely Gamma, Gamma Prime, and Transition-metal rich Carbide. We utilized invariant scalar-based and/or equivariant, tensor-based neural network (NN) approach as implemented in DEEPMD, NEQUIP/ALLEGRO codes respectively. For the training and validation sets, we employed the trajectory results from the ab-initio molecular dynamics (AIMD) and ground state DFT calculations including the energy, force, and virial database from highly diverse compositions, temperatures, and pressures following a "High Entropy Strategy". We systematically developed the Deep Learning potentials for 5, 7, and 10 component systems based on the complexity level of the phase mixtures. To optimize the hyperparameters, we used a series of machine learning (ML) algorithms to lower the RMSE of the force components. We then compare the accuracy of both the potentials developed using the two types of Deep Learning potentials through a variety of large-scale molecular dynamics (MD) simulations. |
Wednesday, March 8, 2023 12:18PM - 12:30PM |
N53.00005: Accelerated Designing of Superhard B-C-O Compounds using Machine Learning and DFT Madhubanti Mukherjee In recent decades, a wide range of applications in defense, mining, manufacturing, and space industries have driven significant interest in designing materials with superior hardness. Compounds containing lighter elements such as boron, carbon, oxygen, and nitrogen are one of the most promising classes of superhard materials due to their ability to form short and covalent bonds. Given the complex procedure, which involves extensive resources and time requirements in both computational and experimental measurements, exploring a large chemical space of such superhard materials has always remained a challenge. Herein, we attempt to accelerate the search for superhard B-C-O compounds by developing machine learning (ML) models for rapid prediction of elastic moduli as proxy properties, followed by subsequent estimation of hardness using Tian’s empirical formula [1]. The ML models have been trained on a set of 10448 compounds with density functional theory (DFT)-computed elastic constants (bulk (K) and shear (G) modulus) using the simple chemical formula derived input space. The developed models exhibit excellent accuracy with R2 of 0.98 and 0.94 for bulk and shear modulus predictions, respectively. The models have further been employed on a set of ternary B-C-O, generated by enumerating a series of BxCyOz compositions with x, y, z ∈ {1, 2, 3, . . . 9}. ML models recommend 335 B-C-O compositions with hardness values of more than 35 GPa among a total of 1320 B-C-O compositions. The predictions include compositions such as B2C3O and B2C5O with hardness values of 36.15 and 55.51 GPa, respectively, consistent with previous findings [2]. Next, the ML predictions have been validated using evolutionary structure predictor and DFT, which identify four unique superhard B-C-O compounds exhibiting hardness ranging from 40 Gpa to 58 Gpa. These structures show mechanical and dynamical stability with relatively lower formation energy, implying a strong possibility of experimental synthesis. |
Wednesday, March 8, 2023 12:30PM - 12:42PM |
N53.00006: Comparing Structural Representations of Grain Boundaries Braxton B Owens The properties of polycrystalline materials are are a function of microstructure. The property-microstructure relationship has motivated many physics-inspired structural representations of materials. Although mainly used to train accurate inexpensive interatomic potentials by means of machine learning, these representations accurately represent some grain boundary properties. Using a database of over 7000 grain boundaries, we evaluate different representations and their abilities to express the relevant structural information. We predict grain boundary energy with each representation in a machine-learned model. Our comparison identifies promising grain boundary descriptors. |
Wednesday, March 8, 2023 12:42PM - 12:54PM |
N53.00007: Benchmarking and Optimization of UF3 Machine Learning Potential on Solids Pawan Prakash, Stephen R Xie, Hendrik Krass, Ajinkya C Hire, Peter Hirschfeld, Matthias Rupp, Richard G Hennig Ab initio calculations offer a promising theory-guided approach to materials discovery and design. Still, ab initio calculations are hindered by their computational cost, limiting the complexity of materials they can evaluate. Machine learning potentials (MLPs) have recently shown their utility in Molecular Dynamic simulations, reaching accuracies comparable to ab initio simulations at a fraction of the computational cost. This work benchmarks the recently published Ultra-Fast Potential (UF3) — which uses linear regression with a cubic B-spline basis to evaluate effective two- and three-body potentials — on elemental systems. We find that UF3 has an accuracy comparable to that of GAP, MTP, NNP(Behler Parrinello), and qSNAP MLPs while being two to three orders of magnitude faster. Apart from speed, another advantage of the UF3 framework is its ability to visualize the learned two- and three-body potentials, helping identify possible "holes" in the learned potential. We have performed extensive hyper-parameter (HP) optimization of UF3. For this, we separate HPs into inner and outer ones according to their computational cost. We then combine grid search on inner HPs with optimization algorithms from the Ray Tune library, leading to accurate and smooth potentials. |
Wednesday, March 8, 2023 12:54PM - 1:06PM |
N53.00008: Evaluation of Thermal Properties of extended 2D Materials using Gaussian Approximation Potentials Alvaro Vazquez-Mayagoitia, Tugbey Kocabas, Cem Sevik, Murat Keceli In recent years, atomically thin two-dimensional materials (2DM) have gained attention due to their flexibility and extraordinary thermal and electronic properties for technological applications. By combining density functional theory (DFT) with Boltzmann transport equation (BTE) it is possible to predict thermal transport properties accurately of these materials; however, the computational cost could be prohibitive for high-throughput calculations or for more realistic simulations with larger super-cell sizes. Herein, we trained Machine Learning potentials, based on Gaussian Approximation Potentials, using an ad-hoc reference dataset of DFT calculations to provide accurate forces for BTE model to estimate the thermal properties of 2DM structures, such as graphene, buckled silicene, and h-XN (X = B, Al, and Ga, as biatomic 2DM) structures. We validated our potentials computing phonon dispersion curves and lattice thermal conductivity via harmonic and anharmonic force constants, respectively, and compared them to DFT results. Additionally, we calculated with out method anharmonic force constants to generate high-order force constants, we found that in the case of 2nd order force constants, GAP predicted not only low-frequency acoustic modes accurately, which are the main heat carriers in semiconductors and insulators, but also relatively high-frequency optical modes. Moreover, this method enabled us to compute 3rd order force constants similar to DFT. |
Wednesday, March 8, 2023 1:06PM - 1:18PM |
N53.00009: Proton dynamics simulations of solid-acid electrolytes using active learning and equivariant neural network force fields. Menghang (David) Wang, Cameron J Owen, Yu Xie, Simon L Batzner, Albert Musaelian, Anders Johansson, Boris Kozinsky Understanding the rate-limiting steps of proton conduction across various solid acid electrolytes and the mechanisms behind superprotonic phase transition is crucial for designing next generation energy fuel cells. Due to the high computational cost of ab-initio molecular dynamics, previous studies of proton dynamics were constrained to a hundred atoms within a few hundred picoseconds and thus led to a limited statistics of proton-hopping events. Previous work has used material-specific empirical force fields to study the phase transition of CsHSO4, but it cannot be generalized to understand the mechanism of other solid acid materials. |
Wednesday, March 8, 2023 1:18PM - 1:30PM |
N53.00010: Polymer Property Prediction via Pre-trained Large Language Model Yuyang Wang, Changwen Xu, Amir Barati Farimani Accurate and efficient evaluation of polymer properties is significant in polymer design. Conventional methods rely on expensive and time-consuming experiments or simulations to assess the material functions. The recent development of Transformer-based large language models has demonstrated superior performance on various applications in natural language processing and computer vision. However, such methods have not been well investigated in polymer science. In this work, we present TransPolymer, a Transformer-based language model built on self-attention for polymer property prediction. We propose a polymer tokenization strategy that encodes material informatics and converts each polymer into a text sequence. Also, TransPolymer benefits from pre-training on large unlabeled data via predicting the masked tokens in a self-supervised learning manner. Experiments have demonstrated that TransPolymer surpasses other baseline machine learning models in various polymer property prediction tasks. Moreover, self-supervised pre-training shows merits over training from the randomly initialized Transformer model. We hope this work provides a promising computational tool for polymer design and understanding structure-property relationships from a data science perspective. |
Wednesday, March 8, 2023 1:30PM - 1:42PM |
N53.00011: Performance Boosting Portable Acceleration of SISSO++ for Symbolic Descriptor Learning Yi Yao, Matthias Scheffler, Christian Carbogno, Luca M Ghiringhelli, Thomas A Purcell Sure-independence screening and sparsifying operator (SISSO) is a powerful, artificial intelligence tool for identifying symbolic descriptors and predictive models [1]. It has been successfully applied for discovering new optimal materials, e.g. [2]. Recently, we have developed SISSO++, a new implementation of SISSO that uses both OpenMP and MPI to approach linear-scaling parallel performance on CPUs [3]. Here, we present an updated algorithm that uses the Kokkos performance-portable programming model to offload the performance-critical region of our algorithm to accelerators, such as Nvidia or AMD GPUs [4]. We demonstrate the performance of these updates by using the prediction of thermal conductivity over rock salts and chalcopyrites as an example and highlight the opportunities opened by the improvement. |
Wednesday, March 8, 2023 1:42PM - 1:54PM |
N53.00012: First-Principles-Informed Machine Learning Study of Defects on the Lithium Metal Surface Hao Yu, Madison Morey, Tianlun Huang, Kubra Cilingr, Ziqing Zhao, Emily Ryan, Brian Kulis, Sahar Sharifzadeh Dendrite formation at the lithium anode-electrolyte interface causes degradation in lithium metal batteries. One factor that influences these branched Li-based growths is the surface and interfacial energetics associated with the presence of defects. These interfaces are disordered on the order of microns and so it is critical to understand the relationship between the microscale energetics and mesoscale. Here, we introduce machine learning (ML) as a way to connect these two scales. As a first step, we develop ML models that can predict the surface energies associated with the introduction of defects into the surface. With a dataset of defect-induced energetics calculated from density functional theory (DFT), we develop various models with conventional ML algorithms and descriptors, and neural networks with defect density features. We evaluate the performance of these models for the data set as well as for individual defect types and densities. |
Wednesday, March 8, 2023 1:54PM - 2:06PM |
N53.00013: Flat bands in full-Heusler crystals – statistical analysis with periodic table deep learning model xiuying zhang Materials with flat bands near or at the Fermi level are a promising route towards strongly correlated states, but the properties demonstration has been severely hindered due to the lack of materials for realization of flat bands. Machine learning is a promising way to speeding up the unearthing of the flat band materials. Here, we propose a convolutional neural network classification model, which only has the periodic table as the input and thus is called periodic table representation (PTR) classifier, to explore flat bands along the high-symmetry paths around the Fermi level and search for the physics behind it. The full-Heusler crystals are chosen as targets because of their abundance in dataset and their much high symmetry with the space group of 225. The PTR model is trained with the full-Heusler crystals in the AFLOW database and then it can accurately classify the flat bands with the AUC of 0.91 for the test set and 0.88 for the full-Heusler crystals in the materials project (MP) dataset. Visualizing the model also gets some interesting phenomina. For example, the model regards the number of valence electrons of the crystal as an important parameter for the flat band prediction: crystals have a high probability to have flat band around the Fermi level if they have valence electrons of about 12×n (n=1,2,3…); crystals have a probability as high as 0.8 to possesse flat bands if their number of unfilled d-orbital electrons is around 6, and the larger number of unfilled p-orbital electrons, the smaller flat band probability. The full-Heusler crystals are reach in magnetic crystals and the PTR can also predict both the flat bands and the ferromagnetic full-Heusler compounds with the accuracy of 0.76. Our work not only provides a quick way to enriching the flat band database, but also provides the specific direction for searing for the flat bands. |
Follow Us |
Engage
Become an APS Member |
My APS
Renew Membership |
Information for |
About APSThe American Physical Society (APS) is a non-profit membership organization working to advance the knowledge of physics. |
© 2024 American Physical Society
| All rights reserved | Terms of Use
| Contact Us
Headquarters
1 Physics Ellipse, College Park, MD 20740-3844
(301) 209-3200
Editorial Office
100 Motor Pkwy, Suite 110, Hauppauge, NY 11788
(631) 591-4000
Office of Public Affairs
529 14th St NW, Suite 1050, Washington, D.C. 20045-2001
(202) 662-8700