Modeling Materials at Realistic time Scales via Optimal Exploitation of Exascale Computers and Artificial Intelligence

A Workshop and Hands-on Tutorial,
July 25 – 29, 2022


The event is planned in terms of two stages: a high-level CECAM workshop and a subsequent hands-on tutorial. Both activities address the concepts and implementations that are needed in order to link the Quantum Mechanical (QM) description of electrons in materials, to the statistical mechanics principles that address the larger time and length scales governing real-life situations.

During the first 3 days, the workshop will focus on recent and important developments addressing exascale scientific computing applications and related artificial intelligence (AI) methods, with a specific focus on urgent and critical aspects in the domain of computational materials science. In particular, we will address how exascale computing can contribute to the enhanced performance of materials modeling, in terms of higher accuracy, precision and degree of inter-operability between different modeling length- and time-scales. These technical aspects will be presented and discussed by leading experts in different domains, thus giving the opportunity to explore similarities and differences in the various current state-of-the-art approaches towards exascale computing, as well as the management of modeling workflows and corresponding output data of interesting materials properties.

Then the following 2 days will consist of tutorials and hands-on demonstrations that will focus on recent progress in (1) first principles simulations and (2) advanced sampling methods and software, and (3) the coupling of first principles molecular dynamics simulations and advanced sampling methods. In particular, examples using the Qbox code coupled with with the SSAGES suite of codes and I-Pi will be discussed in detail, with several hands-on examples.

General Scope

Real materials are not necessarily in thermal equilibrium, and a careful understanding of the micro-structure of any material (e.g. grains and grain boundaries) is crucial for the estimation of important materials properties and functions. Thus, QM techniques have necessarily to be connected to molecular mechanics (MM), large-scale molecular dynamics (MD), kinetic Monte Carlo (kMC), and computational fluid dynamics (CFD), just to name a few methodologies. Very importantly, we need robust connections between all such modelling techniques, including a detailed understanding of the various errors and uncertainties involved. Moreover, in order to properly interpret the corresponding results, we need all such inter-connections to be fully reversible, i.e. not just able to transition from small to large scales but also conversely.

The Handbook of Materials Modeling (2005) is one of the main classical references in this domain of scientific computing [1], and its 2nd edition has since appeared in 2020 [2]. This has now turned into a six-volume major review masterwork, reflecting the significant developments in all aspects pertaining to computational materials research over the past decade or so, including major progress in the formulation of increasingly realistic multi-scale modeling approaches, workflows and models. However, two recent innovations in materials modeling applications are still relatively poorly and sparsely covered in the currently available review literature, namely exascale computing and related artificial intelligence (AI)-based methods. These two topics will be the main focus of our attention within our proposed workshop and associated tutorial school.

[1] Handbook of Materials Modeling, 2005, S. Yip (ed), ISBN 978-1402032875, Springer, Cham
[2] Handbook of Materials Modeling, 2nd ed., 2020, W. Andreoni and S. Yip (eds), ISBN 978-3319788760, Springer, Cham


The event will start with a high-level workshop (3 days). Talks, discussions, and poster sessions will be held as a regular meeting involving the physical presence of all participants (some but very few talks may be presented virtually).

Each of the five main sessions during the first part of the workshop (throughout the first three days) will start with an introduction (15 minutes) by a renowned scientist, the so-called “moderator” for that particular session. The subsequent talks in the corresponding session will then last for 30 minutes each, and will be followed in turn by 10 minutes of general Q&A discussion.

The following 2-day hands-on tutorial  will include coupling first principles molecular dynamics simulations and advanced sampling methods using the Qbox code coupled with with the SSAGES suite of codes and I-Pi. Several examples will be discussed with hands-on demonstrations, and opportunities will be provided for students to develop simulation strategies of direct relevance to their own research with the help of expert instructors.

Date & Location

The Workshop will take place during the period of July 25-29, 2022, at the Humboldt Universität zu Berlin at Campus Adlershof.

A more detailed description of way can be found here.

Hotel Accommodation

Accommodation will be at nearby hotel: Dorint Berlin-Adlershof


Programme (Download as PDF, 50kb)

Room 0.119, Erwin-Schrödinger-Zentrum, Rudower Chaussee 26, 12489 Berlin (GoogleMaps, OpenStreetMaps)

14:45 Welcome Matthias Scheffler IRIS Adlershof / Humboldt-Universität zu Berlin & Fritz Haber Institute of the Max Planck Society, Germany

Session 1: Architectures of exascale computers and necessary coding concepts
Moderator: Erwin Laure

15:15 Introduction Erwin Laure Max Planck Computing and Data Facility, Germany
15:30 The supercomputer Fugaku - AI and Big Data Kento Sato RIKEN Center for Computational Science, Japan
16:10 Coffee break
16:30 El Capitan: The LLNL Exascale System Bronis de Supinski Lawrence Livermore National Laboratory, USA
17:10 Material modelling in Aerospace Applications using HPC and Cloud computing Carlo Cavazzoni Leonardo Lab, Italy
17:50 Risc-V Vector CoDesign in EPI Jesus Labarta Mancho Barcelona Supercomputing Center - Centro Nacional de Supercomputación (BSC-CNS), Spain
All talks 30 minutes + 10 minutes discussion
Dinner with Poster Session
Foyer, IRIS Adlershof,
Zum Großen Windkanal 2, 12489 Berlin, Germany (GoogleMaps,OpenStreetMaps)

Session 2: Multi-scale modeling at the exascale
Room 0.119, Erwin-Schrödinger-Zentrum, Rudower Chaussee 26, 12489 Berlin (GoogleMaps, OpenStreetMaps)
Moderator: Sara Bonella

09:00 Introduction Sara Bonella CECAM / École Polytechnique Fédérale de Lausanne, Switzerland
09:15 Exascale challenge for workflows and multi-scale modeling Modesto Orozco Institut de Recerca Biomèdica, IRB Barcelona, Spain
09:55 Exascale workflows Geoffroy Hautier Dartmouth College, USA
10:35 Coffee break
10:55 Mesoscale algorithms and codes in exascale architectures Massimo Bernaschi Consiglio Nazionale delle Ricerche (CNR), Italy
12:15 Lunch break at I Due Amici (GoogleMaps, OpenStreetMaps)
13:45 Kinetic Monte Carlo Qian Yang University of Connecticut, USA
14:25 Multiscale modeling in soft and biological matter Matej Praprotnik Kemijski Inštitut, National Institute of Chemistry, Slovenia
15:05 Coffee Break
All talks are 30 minutes + 10 minutes discussion

Session 3: AI for molecular modelling
Room 0.119, Erwin-Schrödinger-Zentrum, Rudower Chaussee 26, 12489 Berlin (GoogleMaps, OpenStreetMaps)
Moderator: Kurt Kremer

15:25 Introduction Kurt Kremer Max Planck Institute for Polymer Research, Germany
15:40 Compound discovery by multiscale modeling and machine learning Tristan Bereau Independent
16:20 Machine Learning Dielectric Screening for the Simulation of Excited State Properties of Molecules and Materials Giulia Galli Universität of Chicago, USA
17:00 Extracting Design Principles from Physics-Inspired Machine Learning Rose K.
University of Wisconsin, Madison USA (soon)
17:40 Big data science in porous materials Berend Smit École Polytechnique Fédérale de Lausanne, Switzerland
18:20 End
All talks are 30 minutes + 10 minutes discussion
Dinner with Poster Session
Foyer, IRIS Adlershof,
Zum Großen Windkanal 2, 12489 Berlin, Germany (GoogleMaps, OpenStreetMaps)

Session 4: Artificial intelligence concepts
Room 0.119, Erwin-Schrödinger-Zentrum, Rudower Chaussee 26, 12489 Berlin (GoogleMaps, OpenStreetMaps)
Moderator: Claudia Draxl

09:00 Introduction Claudia Draxl IRIS Adlershof / Humboldt-Universität zu Berlin & Fritz Haber Institute of the Max Planck Society, Germany
09:15 Finding structure in data, identifying maps of materials properties, and detecting the "materials genes" Matthias Scheffler IRIS Adlershof / Humboldt-Universität zu Berlin & Fritz Haber Institute of the Max Planck Society, Germany
09:55 Molecular Dynamics wih the Deep Potential Method Roberto Car Princeton University, USA
10:35 Coffee break
10:55 Machine learned potential-energy surfaces Cecilia Clementi Freie Universität Berlin, Germany
11:35 Data-driven discovery of rare phenomena Mario Boley Monash University, Australia
12:15 Lunch break at I Due Amici (GMaps, OSM)
All talks are 30 minutes + 10 minutes discussion

Session 5:  Challenges in atomistic modelling
Room 0.119, Erwin-Schrödinger-Zentrum, Rudower Chaussee 26, 12489 Berlin (GoogleMaps, OpenStreetMaps)
Moderator: Ignacio Pagonabarraga

13:45 Introduction Ignacio Pagonabarraga CECAM / École Polytechnique Fédérale de Lausanne, Switzerland
14:00 Exhilarating exascale explorations in materials space Nicola Marzari École Polytechnique Fédérale de Lausanne, Switzerland
14:40 Design of macromolecular products and processes from scratch Juan de Pablo University of Chicago, USA
15:20 Grand canonical replica exchange MD from first principles Luca Ghiringhelli IRIS Adlershof / Humboldt-Universität zu Berlin & Fritz Haber Institute of the Max Planck Society, Germany
16:00 Coffee break
All talks are 30 minutes + 10 minutes discussion
16:30 Departure at Hotel Dorint
Conference outing: Graffiti tour at the former US listening station
Conference dinner in the Grunewald

Hands-on tutorial: Day 1
Room 1'427, Department of Physics, Lise Meitner-Haus, Newtonstraße 15 (GoogleMaps, OpenStreetMaps)

9:00 Introduction to ab initio molecular dynamics and DFT

Giulia Galli, University of Chicago, Pritzker School of Molecular Engineering, USA
Francois Gygi, University of California Davis, USA
11:00 Coffee break at IRIS Adlershof
11:15 Introduction to advanced sampling methods

Juan de Pablo, University of Chicago, Pritzker School of Molecular Engineering, USA
Ludwig Schneider, University of Chicago, Pritzker School of Molecular Engineering, USA
13:15 Lunch break at IRIS Adlershof
14:30 Hands-on: Introduction to QBox

Francois Gygi, University of California Davis, USA
Arpan Kundu, University of Chicago, Pritzker School of Molecular Engineering, USA
15:30 Hands-on: Introduction to SSAGES

Pablo Zubieta, University of Chicago, Pritzker School of Molecular Engineering, USA
Ludwig Schneider, University of Chicago, Pritzker School of Molecular Engineering, USA
16:30 Coffee break at IRIS Adlershof
16:45 Hands-on: Coupling of ab initio and advanced sampling techniques

Elizabeth M.Y. Lee, University of Chicago, Pritzker School of Molecular Engineering, USA
Gustavo Perez, University of Chicago, Pritzker School of Molecular Engineering, USA
18:00 BBQ @ Gerdans, Erwin Schrödinger-Zentrum

Hands-on tutorial: Day 2
Room 1'427, Department of Physics, Lise Meitner-Haus, Newtonstraße 15 (GoogleMaps, OpenStreetMaps)

9:00 Hands-on Problem # 1

Arpan Kundu, University of Chicago, Pritzker School of Molecular Engineering, USA

Elizabeth M. Y. Lee, University of Chicago, Pritzker School of Molecular Engineering, USA
11:00 Coffee break at IRIS Adlershof
11:15 Hands-on Problem # 2

Gustavo Perez, University of Chicago, Pritzker School of Molecular Engineering, USA
Pablo Zubieta, University of Chicago, Pritzker School of Molecular Engineering, USA
Ludwig Schneider, University of Chicago, Pritzker School of Molecular Engineering, USA
13:15 Lunch at IRIS Adlershof
14:30 Parallel Sessions

Giulia Galli, Francois Gygi, Juan de Pablo, and all PD and students of previous sessions
Session 1: Hands-on Problem # 3 Session 2 : Exploratory/discovery session with experts
16:30 End

Poster Abstracts (Download as PDF, 135kb)

Ultrahigh Out-of-Plane Piezo­electricity in 2D monolayers

Raihan Ahammed (raihan.ph18203(ät) and Abir De Sarkar
Institute of Nano Science and Technology,
Knowledge City, Punjab, India

The simultaneous occurrence of gigantic piezoelectricity and Rashba effect in two-dimensional materials are unusually scarce. Inversion symmetry occurring in MX3 (M=Ti, Zr, Hf; X= S, Se) monolayers is broken upon constructing their Janus monolayer structures MX2Y (X≠Y=S, Se), thereby inducing a large out-of-plane piezoelectric constant, d33 (~68 pm/V) in them. d33 can be further enhanced to a super high value of ~1000 pm/V upon applying vertical compressive strain in the van der Waals bilayers constituted by interfacing these Janus monolayers. Therefore, d33 in these Janus transition metal tri-chalcogenide bilayers reach more than four-fold times that of bulk ceramic PZT material (~268 pm/V). The absence of horizontal mirror symmetry and the presence of strong spin-orbit coupling (SOC) cause Rashba spin splitting in electronic bands in these Janus 2D monolayers, which shows up as an ultrahigh Rashba parameter, αR ~ 1.1 eVÅ. It can be raised to 1.41 eVÅ via compressive strain. Most of the 2D materials reported to date mainly show in-plane electric polarization, which severely limit their prospects in piezotronic devices. In this present work, the piezoelectricity shown by the Janus monolayers of Group IV transition metal tri-chalcogenides and their bilayers is significantly higher than the ones generally utilized in the form of three-dimensional bulk piezoelectric solids, e.g., α-quartz (d11 = 2.3 pm/V), wurtzite-GaN (d33 = 3.1 pm/V), wurtzite-AlN (d33 = 5.6 pm/V). It is exceedingly higher than that in Janus multilayer/bulk structures of Mo and W based transition metal dichalcogenides, e.g., MoSTe (d33~10 pm/V). The 2D Janus transition metal trichalcogenide monolayers and their bilayers reported herewith straddle giant Rashba spin splitting and ultrahigh piezoelectricity, thereby making them immensely promising candidates in the next generation electronics, piezotronics, spintronics, flexible electronics and piezoelectric devices. We have also performed a high-throughput computational screening for two-dimensional (2D) piezoelectric materials. From the existed 2D materials database, we have screened 252 materials which show piezoelectricity out of 1000 stable 2D materials.

R. Ahammed, N. Jena, A. Rawat, M.K. Mohanta, Dimple, and A.­ De Sarkar
Ultrahigh Out-of-Plane Piezoelectricity Meets Giant Rashba Effect in 2D Janus Monolayers and Bilayers of Group IV Transition-Metal Trichalcogenides
J. Phys. Chem. C 124, 21250-21260 (2020)

Machine Learning-Enabled Prediction of Electronic Properties of Radical- Containing Polymers at Coarse-Grained Resolutions

Riccardo Alessandri (alessandri(ät) and Juan J. de Pablo
Pritzker School of Molecular Engineering, University of Chicago, USA

Non-conjugated radical-containing polymers are a class of charge-carrying polymers that rely on pendant stable radical sites to transport charges successfully. They constitute promising materials for applications in all-organic energy storage or memory devices. The properties of these and other soft electronic materials depend on the coupling of electronic and conformational degrees of freedom over a wide range of spatiotemporal scales. Predictive modeling of such properties requires multiscale approaches that efficiently connect quantum-chemical calculations to mesoscale coarse-grained (CG) methodologies.
We present an efficient computational scheme that leverages supervised machine learning (ML) to predict electronic-structure information pertaining to charge transport at CG resolutions. ML models are trained on data coming from quantum-chemical calculations on all-atom conformations sampled from condensed-phase simulations. Predictions of electronic couplings, spin densities, and energy levels are studied as a function of CG mapping and resolution. Trained ML models subsequently enable electronic property predictions directly from CG polymer morphologies. We evaluate both systematic and building-block coarse-graining techniques. We validate the approach by comparing to the standard methodology that requires backmapping to atomistic resolution and subsequent quantum-chemical calculations. As such, the proposed ML- assisted scheme drastically accelerates multiscale computational workflows that connect electronic properties to mesoscale morphological features, thereby enabling high-throughput modeling aimed at the understanding and rational design of radical-containing polymers and other soft electronic materials.


Alexander Buccheri (abuccheri(ät)
Humboldt-Universität zu Berlin, Germany

The GW approximation has become a popular method for predicting quasiparticle excitations in both solids and molecules alike, due to its high accuracy with moderate computational effort [1, 2]. Direct implementation of the GW method scales as O(N4) with system size, where both the dielectric function and the correlation self-energy are formulated in terms of a product basis of Kohn-Sham wave functions [3]. We propose an efficient LAPW implementation of GW in the all-electron code, exciting [4], based on the space-time method [5], which will reduce the scaling of exciting’s G0W0 method to O(N3).
This work is part of a larger effort towards developing an excited-state library, GreenX, under the NOMAD CoE. Like other libraries in the CoE, this will be a mix of code-agnostic implementations, which can be shared between code families with different basis-set types, and code-specific features. Abstraction of the most computationally-intensive routines required for excited-state calculations into an efficient library, and the pooling of developer resources through the CoE, underpins the broader ambition of migrating the code families towards excited-states simulations on exascale hardware.

[1] Golze, Dorothea, Marc Dvorak, and Patrick Rinke. „The GW compendium: A practical guide to theoretical photoemission spectroscopy.“ Frontiers in chemistry 7 (2019): 377.
[2] van Setten, Michiel J., et al. „GW 100: Benchmarking G0W0 for molecular systems.“ Journal of chemical theory and computation 11.12 (2015): 5665-5687.
[3] Jiang, Hong, et al. „FHI-gap: A GW code based on the all-electron augmented plane wave method.“ Computer Physics Communications 184.2 (2013): 348-366.
[4] Gulans, Andris, et al. „exciting: a full-potential all-electron package implementing density- functional theory and many-body perturbation theory.“ Journal of Physics: Condensed Matter 26.36 (2014): 363202.
[5] Kutepov, Andrey L., Viktor S. Oudovenko, and Gabriel Kotliar. „Linearized self-consistent quasiparticle GW method: Application to semiconductors and simple metals.“ Computer Physics Communications 219 (2017): 407-414.

Denoising Autoencoder Trained on Simulation-Derived Structures Reveals Tetranucleosome Motifs in STEM Images of Chromatin

Walter Alvarado (walt(ät)
Pritzker School of Molecular Engineering, University of Chicago, USA

Advances in scanning electron microscopy have led to an increase in the quantity and quality of imaging data. While existing algorithms can determine descriptive physical parameters, resolvability remains a challenge for many of these methods. Therefore, there is a need for data-driven approaches that reduce noise and allow for accurate structural predictions from experimental images. ChromSTEM, a combination of DNA-specific staining in scanning transmission electron microscopy, as allowed for the 3D study of genome organization. By leveraging convolutional neural networks and molecular dynamics simulations, we have developed a denoising autoencoder (DAE) capable of providing nucleosome-level resolution. Our DAE is trained on synthetic images generated from simulations of the chromatin fiber using the 1CPN model of chromatin. We find that our DAE is capable of removing noise commonly found in high angle annular dark field (HAADF) STEM experiments and is able to learn structural features driven by the physics of chromatin folding. We find that our DAE outperforms other well-known denoising algorithms without degradation of structural features. In addition, we are able to identify several tetranucleosome motifs and observe the absence of a 30-nm fiber, which has been suggested to serve as the higher-order structure of the chromatin fiber.

Thermodynamics of Ionic Bon­ding with Crown Ethers in Aqueous Solution

Ramón González-Pérez (rgonza22(ät) and Jonathan K. Whitmer
University of Notre Dame, Indiana, USA

The usage of Lithium for powering consumer electronics has increased its extraction from the land. However, the extraction process is complicated and polluting, and there- fore, it has encouraged interest in Lithium recycling methods. In this work, Molecular simulations of aqueous solutions of three different crown ethers (12-Crown-4, 15-Crown-5 and 18-Crown-6) have been performed using Molecular Dynamics to explore their binding processes to Li+. Free energy calculations with umbrella sampling were used to compute the potential of mean force PMF of each system. The simulations were carried out on GROMACS using the GAFF force field and the explicit-solvent model TIP3P for water. In addition, the binding processes of the crown ethers to Na+ were also examined for com- parison.

Infrared-active acoustic-optical phonon modes in two dimensional organic/inorganic interfaces

Hanen Hamdi (mailhanenhamdi(ät), Jannis Krumland, and Caterina Cocchi
Carl von Ossietzky Universität Oldenburg

The growing interest in hybrid inorganic/organic interfaces formed by conjugated molecules adsorbed on transition-metal dichalcogenide monolayers has stimulated a number recent studies on the electronic properties of these compounds¹,²,³. However, little is known about the phonon characteristics of these hybrid materials, which are crucial for a full understanding of their photophysical properties. In this work, we use first-principles calculations based on density
-functional (perturbation) theory, to study the hybrid system formed by the pyrene molecule adsorbed on either MoS2 or MoSe2 monolayers. We find that, in these hybrid materials, new phonon modes are activated, compared to those of the individual constituents and play a crucial role to determine some of the main prominent optical peaks in the infrared spectra. Furthermore, we analyze the coupling modes that appear due to the hybridisation.

[1] F. Mahrouche, K. Rezouali, S. Mahtout, F. Zaabar, and A. Molina Sánchez, Phonons in WSe2/MoSe2 van der waals heterobilayers, physica status solidi (b) 259, 2100321.
[2] D. Sercombe, S. Schwarz, O.D. Pozo-Zamudio, F. Liu, B.J. Robinson, E.A. Chekhovich, I.I. Tartakovskii, O. Kolosov, and A.I. Tartakovskii, Optical investigation of the natural electron doping in thin MoS2 films deposited on dielectric substrates, Scientic Reports (2013).
[3] Jannis Krumland and Caterina Cocchi, IOP 2021 Electron. Struct. 3 044003

Long Range Interactions in Atomistic Machine Learning

Kevin Kazuki Huguenin-Dumittan (kevin.huguenin-dumittan(ät)
École Polytechnique Fédérale de Lausanne, Switzerland

Machine learning (ML) based methods to study materials at the atomistic scale have experienced a surge in interest in the last couple of years. One central assumption in many models for atomistic simulations is locality: the behavior of an atom primarily depends on its neighborhood. This is the assumption that leads to more efficient algorithms, making the methods applicable to large systems that could only be studied using classical physics, as opposed to quantum mechanical approaches like density functional theory, effectively bridging the two worlds.
Despite of the successes, effects including Debye-Hückel screening or nonlocal charge transfers cannot be captured in this way. The long distance equivariant (LODE) representation, first introduced in 2019, allows one to create ML models that go beyond this approximation. The approach as well as its implementation are strongly inspired by the Ewald based techniques in conventional molecular dynamics (MD).
We will present an overview of recent developments on how to incorporate long range interactions in atomistic ML models. Starting from extensions of the original LODE framework, we will demonstrate preliminary results showing the use of these descriptors, as well as the challenges that arise when combining this approach with traditional MD.

Ab-initio Prediction of Adsorption Isotherms of Water with a Metal- Organic Framework

Nicole Mancini (nicole.mancini(ät)
Humboldt-Universität zu Berlin, Germany

The interaction of water with the metal organic framework Mg2(2,5-dioxido-1,4-benzenedicarb-oxylate) (MOF-74-Mg), will be studied using chemically accurate quantum chemical methods. Adsorption
structures with one, two and five H2O molecules per Mg2+ site will be investigated. The calculated isotherms will be compared with experimental data.
Subsequently, the formation of defects in the crystalline structure by hydrolysis will be studied to understand the wide variations of experimental isotherms. The calculations will use a hybrid QM:QM approach, which provides coupled cluster corrections to density functional theory and has been shown to yield chemical accuracy for molecule-surface interactions.

Empowering Carbon & Energy Catalysis with HPC, Data, AI and the Cloud

Vivek Sinha (vivek.sinha(ät) and Farnaz Sotoodeh
Carbon and energy catalysis (C2CAT), Lisserbroek, Netherlands

Catalytic processes and materials play a key role in enabling power-2-X (P2X) transformations. C2CAT specializes in the design of customized catalysts for crucial P2X transformations such as the electrocatalytic nitrogen reduction reaction and conversion of CO2, using proprietary technology which boosts the catalytic activity. Our approach is rooted in fundamental research on design and discovery of new materials using state-of-the-art computational modelling and simulations techniques. We implement and commercialize these multi-scale discoveries in real-world applications through synthesis of novel materials and their scale-up via various unique technologies.
This poster will discuss some of the key P2X applications that leverage from multiscale modelling of catalysis combining quantum chemistry with microkinetic modelling, empowered via artificial intelligence (AI), and high- performance computing (HPC) powered over the cloud.

High-throughput ab-initio calculations as a tool for screening tabulated experimental data

Daria M. Tomecka¹ (tomeckadm(ät), S. Cottenier¹,², V. Van Speybroeck¹,², and M. Waroquier¹
¹ Center for Molecular Modeling, Ghent University, Belgium
² Department of Materials Science and Engineering, ebd.

In this contribution we present calculated band gaps for a set of ca. 250 semiconductors for which experimental band gaps are available in standard tabulations [1]. Calculations were performed with the LAPW code WIEN2k [2], using the PBE and modified Becke-Johnson (mBJ) functional. Correlations between the two sets of calculations and experiment are presented and discussed, as well as correlations between the two sets of calculations themselves. It is shown to which extent such an approach can be used to identify “suspicious” entries in tabulated experimental data.

[1] CRC Handbook of Chemistry and Physics, 91st Edition, CRC Press, William M. Haynes, National Institute of Standards and Technology, Boulder, Colorado, USA
[2] P. Blaha et al., WIEN2K, An Augmented Plane Wave + Local Orbitals Program for Calculating Crystal Properties (Karlheinz Schwarz, Techn. Universitaet Wien, Austria), 1999. ISBN 3-9501031-1-2

Uncertainty Modelling for Property Prediction of Double Perovskites

Simon Teshuva (simon.teshuva(ät)
Monash University, Victoria, Australia

Statistical predictive models for double perovskite properties are of high interest, because the perovskite structure allows relatively accurate property prediction and at the same time provides enough flexibility to yield a huge number of different materials of which some are likely relevant for important applications. While promising performance results have been published for this class of materials, they typically refer only to the predictive performance as, e.g., measured by the root mean squared error. However, active learning strategies for effective materials screening rely critically not only on accurate predictions but also on adequate uncertainty estimates as provided by probabilistic models.
Here, we study the predictive performance of two popular machine learning models, Gaussian processes and random forests, together with the quality of their uncertainty estimates. This study is based on a dataset of over 800 single (ABO_3) and double (A_2BB‘O_6) cubic perovskite oxides with computed bulk modulus, cohesive energy, and bandgap. As a first observation, we show that random forests outperform Gaussian processes in terms of predictive performance for all three properties.
Moreover, we show that Gaussian processes, while providing sound Bayesian uncertainty estimates, can have inferior performance when their assumption of isometric smoothness of the target property is not met. In this case, as exemplified by the double perovskite bandgaps, random forests provide a better alternative, despite their rather ad-hoc uncertainty estimates. Improving these estimates thus appears to be a promising direction for future research, potentially leading to substantially improved procedures for computationally efficient sequential candidate screening.

Periodic Coupled-cluster Theory and Applications: Interface to FHI-aims and VASP

Evgeny Moerman (moerman(ät)
The NOMAD Laboratory, Fritz-Haber-Institut der MPG, Germany

Coupled-cluster (CC) theory achieves a high level of accuracy for a wide range of properties in atoms and molecules at affordable computational cost and is generally considered the "gold-standard" in quantum chemistry. However, only few codes exist for CC calculations of periodic systems. The Cc4s code[1] allows for periodic CCSD(T)-level calculations and it has been interfaced to VASP[2] and, recently, to FHI-aims[3, 4]. Together these interfaces, which cover atom-centered basis and plane-wave codes, can be generalized to interface most other ab initio software packages to Cc4s. This enables a wide range of applications ranging from molecular systems in the gas phase to solids and surfaces. The numerical heavy lifting in Cc4s is outsourced to a tensor contraction framework, which currently is the Cyclops Tensor Framework[5]. However, a generalized interface has been implemented, so that in principle any parallel tensor contraction library can be used by Cc4s. Using this strategy, it has been possible to compute systems with over 300 electrons and about 2000 virtual states. With the exploitation of the block-sparsity of tensors in periodic Coupled-cluster theory, which is currently being implemented in Cc4s, it is expected that conventional Coupled-cluster calculations with up to a 1000 electrons will become accessible.

[1] The Cc4s web page.
[2] Kresse et al. (1996). “Efficiency of ab-initio total energy calculations for metals and semiconductors using a plane-wave basis set”. Computational materials science 6 (1), 15–50. .
[3] Moerman et al. (2022). “Interface to high-performance periodic coupled-cluster theory calculations with atom-centered, localized basis functions”. Journal of Open Source Software 7 (74), 4040. .
[4] The FHI-aims web page. Accessed: 11-05-2022.
[5] Solomonik et al. (2014). “A Massively Parallel Tensor Contraction Framework for Coupled-Cluster Computations”. Journal of Parallel and Distributed Computing 74 (12), 3176–3190. .

AI with Experimental and Theoretical Data toward the Understanding CO2 Hydrogenation Catalysis: The Role of the Support Materials

Ray Miyazaki¹ (miyazaki(ät), Kendra Belthle², Harun Tuysuz², Lucas Foppa¹, and Matthias Scheffler¹

¹ The NOMAD Laboratory, Fritz-Haber-Institut der MPG, Germany
² Max-Planck-Institut fur Kohlenforschung, Germany
The genesis of organic molecules from CO2 at a hydrothermal vent, which is a fissure on the seafloor, is one of the theories for the origin of life [1]. We focus on CO2 hydrogenation catalyzed by cobalt nanoparticles supported on SiO2, which mimic the environment of a hydrothermal vent. In particular, we investigate the role of support materials by using several amorphous SiO2 supports incorporating different elements (e.g., Ti, Zr). In this study, the experimental selectivity toward organic molecules (e.g., methanol, formic acid) is modeled by the sure-independence screening and sparsifying operator (SISSO) AI approach [2]. Our approach identifies the key descriptive parameters correlated to the selectivity, which lead to a better understanding of the origin of life and to design of novel CO2 hydrogenation catalysts.

[1] M. Preiner et al., Nat. Ecol. Evol., 4, 534-542 (2020).
[2] R. Ouyang et al., Phys. Rev. Mater., 2, 083802 (2018)

Hydrogen adsorption on Pd surfaces and its effect on CO2 activation

Herzain Rivera (rarrieta(ät)
The NOMAD Laboratory, Fritz-Haber-Institut der MPG, Germany

An accurate description of the surface of Pd-based catalysts under reaction conditions is a critical step toward a deeper understanding of catalyst reactivity. Herein, by modeling the phase diagram of the (111) and (100) surfaces of face-centered cubic Pd via ab initio atomistic thermodynamics, we predict the stable hydrogen coverages for a wide range of temperatures and H2 pressures. The hydrogen coverage at the experimental conditions used for CO2 hydrogenation plays a major role in the reactivity, as it hinders the chemisorption of activated CO2. The calculated data will be the basis for subsequent subgroup-discovery analysis on CO2 activation.

Surface reconstructions and diffusion properties of beta-Ga2O3 surfaces from first principles

Konstantin Lion (lion(ät) and Quaem Hassanzada
The NOMAD Laboratory, Fritz-Haber-Institut der MPG, Germany

This project is part of the Berlin-centered Leibniz ScienceCampus “Growth and Fundamentals of Oxides (GraFOx) for electronic applications” which combines the expertise of its eight partner institutions in the field of oxide research. The primary focus is on synthesizing oxides to the highest material quality and thoroughly studying them in terms of their surface structure, microstructure, and optical as well as optoelectronic properties.
The transparent conducting oxide Ga2O3, exhibiting a band gap of about 4.9eV, is a very promising candidate for several applications, such as semiconducting lasers and transparent electrodes for UV optoelectronic devices and solar cells. As such, the bulk properties of its thermodynamically stable β phase have been extensively studied in the last two decades. The surface properties, however, playing a vital role in epitaxial growth, electrical contacts, and gas sensors are still not well understood.
In this project, we study the reconstructions of β-Ga2O3(001) in realistic conditions from first principles. With the replica-exchange grand-canonical (REGC) approach [1], we screen for possible surface reconstructions in a reactive oxygen atmosphere in an unbiased way which includes vibrational contributions with full anharmonicity. The obtained metastable structures are used to construct a surface phase diagram using conventional ab initio atomistic thermodynamics (aiAT). By combining the two approaches, we can identify two novel 1x2 reconstructions that exhibit a region of stability in the phase diagram. Our structural model is consistent with STEM images of homoepitaxially-grown (001) films. We also illustrate our current work on the diffusion of gallium and oxygen atoms on the (001) surface. We outline how we will combine the obtained energy barriers, machine learning, and kinetic Monte Carlo methods to model the growth of group-III sesquioxides alloys by molecular beam epitaxy.

[1] Y. Zhou, M. Scheffler, L. M. Ghiringhelli, Phys. Rev. B 100, 174106 (2019).


For all further information related to this event, please send an email to the following address:

Phone: +49-30-2093-66350
Fax: +49-30-2093-13-66350

Postal address:
Humboldt-Universität zu Berlin
IRIS Adlershof
Sitz: Zum Großen Windkanal 2

Unter den Linden 6
10099 Berlin, Germany