From predictive modelling to machine learning as versatile tools for materials design

Resume : Organic semiconductors are promising materials for cheap, scalable and sustainable electronics, light-emitting diodes and photovoltaics [1,2]. The challenge is to find molecules with the right properties in the vast chemical compound space. For example, the molecules should be durable, non-toxic, and easy to process. For organic photovoltaic cells, the ionization energy should fit to the optical spectrum of sun light. Due to the sheer number of conceivable compounds, it is prohibitively expensive to determine the properties of each compound with ab-initio methods. This is where machine-learning steps in. Recently, a variety of graph networks have been developed that can be trained to predict molecular properties, e.g. the binding energy, from the molecular structure. Here, we build upon the physics-motivated SchNet model [3], which is based on pairwise interactions in 3D space and respects all relevant symmetries. The SchNet model sums up atomic contributions to arrive at the total energy, which is the natural approach for extensive quantities. For organic photovoltaics, we are interested in more complex quantities like the HOMO and LUMO energies of a molecule. Therefore, we extend SchNet with the Set2Set unit [Vin16], which has more expressive power than a plain sum. Most previous models have been trained and evaluated on rather small molecules (in particular the "QM9" dataset). Here, we extend the scope of these machine learning-methods by adding also larger molecules from other sources and establish a consistent train-validation-test split. Both contributions improve the accuracy of graph networks to predict properties relevant for organic photovoltaics and other applications. Our model will be used in conjunction with geometry-generating methods and genetic algorithms [Hen20,Arap18] to discover better material. [1] Cheng and Yang, Accounts of Chemical Research 53, 1218 (2020). [2] Gaul et al., Nature Materials 17, 439 (2018). [3] Schütt et al., J. Chem. Phys. 148, 241722 (2018). [4] Vinyals, Bengio & Kudlur, arXiv:1511.06391 (2015). [5] Henault, Rasmussen, & Jensen, PeerJ Physical Chemistry 2, e11 (2020). [6] Arapan, Nieves & Cuesta-López, J. Appl. Phys. 123, 083904 (2018). This work is supported by the European Union's Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 883256.

Resume : Cylindrical pressure vessel should stand very high pressure, for which carbon fibers are wound on it to strengthen pressure vessel’s properties. Most of research is related to find the burst pressure of vessels to pursue optimal vessel design and optimal pattern of wounded filaments. This work is concerned with optimal patterns. Filament-wounded patterns consist of thickness, orientation and stacking sequence, determining the burst pressure. Usually, the number of patterns available for pressure vessel are restricted because filament winding process (helical, hoop and axial winding) constrains the angle and sequence. In this study, we focused on finding optimal pattern using mechanical simulation and artificial intelligence. For a given vessel design, the vessel was divided into three parts (top, dome, and cylinder sections). We defined the optimal pattern as having same performance to withstand burst pressure with less carbon filament. Then, various simulations were performed using different filament geometries to find design rules which can be applied to all vessel designs (details will be presented at the Conference). Next, we studied an optimum pattern and vessel geometry for a given pressure vessel using a neural network built to find the burst pressure based on pattern and vessel geometry. Finally, the inverse step of this neural network was developed to provide an optimal pattern and vessel geometry when a given burst pressure is satisfied.

Resume : Gold nanoparticles show interesting catalytic and optical properties and have been used in a plethora of applications ranging from energy harvesting to biomedical. These properties are greatly affected by the shape of the nanoparticle, hence it would be of great importance to be able to predict their shape. Typically, shape is predicted theoretically by means of the Wulff construction that only takes into account the surface energy. This method is excellent for large nanoparticles, but has limited accuracy for smaller nanoparticles in which the contributions from the edges become significant. Moreover, there is no consensus in the literature on the values of edge energies of common metals. In our work, we focus on the calculation of the edge energies of gold nanoparticles using a data-informed machine-learning (ML), multiple linear regression algorithm. The algorithm was provided with structural data for the nanoparticles along with the total energy as was calculated using either interatomic potentials or Density Functional Theory (DFT) and fitted the data to a relatively simple model. The latter considers the total energy of a nanoparticle to be a sum of contributions from bulk, surface, edge and vertex atoms. From the ML fit, the average energy of atoms based on their geometric position is directly extracted from the calculated coefficients of the model. Leave-one-out cross validation yielded excellent predictions for the total energy of the nanoparticles with a mean absolute percentage error of 1 meV. The present method reproduces correct bulk- and surface-energies of gold for DFT as well as several different interatomic potentials. The predicted edge energies are well converged with respect to the sample size. This research work was supported by the Hellenic Foundation for Research and Innovation (HFRI) under project "MULTIGOLD" (HFRI-FM17-1303, KA 10480)

Resume : Amorphous chalcogenide systems such as GeSe2, can be considered prototypical binary chalcogen-based disordered materials and a nowadays appealing material for advanced memories, optical switches for neuromorphic, quantum computing and optical interconnects. However, key questions remain unanswered about the structural features characterizing chalcogenide materials, in particular for what concern the amorphous phase. This study focuses on the set-up of a computational procedure combining first-principles molecular dynamics (FPMD) and machine-learning interatomic potential (MLP) MD to quantitatively assess the structural properties of a prototypical chalcogenide liquid and glass. We first produce structural models of GeSe2 liquid and glass using first-principles molecular dynamics (FPMD) in quantitative agreement with available experimental data. Secondly, we use FPMD energies, forces and virials to develop through a Gaussian Process Regression (GPR) kernel-based scheme a Gaussian Approximation Potential-type of machine learning potential (GAP, MLP) for modelling, by classical MD, GeSe2 liquid and glass.

Resume : Light emitting diodes (LED) have increasingly gained attention for being a good candidate to substitute the outdated incandescent, halogen, and compact fluorescent light bulbs. The reason for the increased popularity of LEDs is their energy efficiency. While LEDs greatly energetically outperform other light sources, their production has a severus ecological impact related to extraction and refining of rare-elements used in their fabrication. Recently the concept of bio-LEDs was introduced to substitute classical LEDs. Bio-LEDs are an innovative technology based on fluorescent proteins such as the green fluorescent protein (GFP). Fluorescent proteins can be obtained from bacteria reducing to a large extent the ecological impact. However, they are prone to denaturation in non-physiological conditions, which limits their use as active materials in optoelectronic applications. To overcome this issue, in Bio-Led they are placed in a polymer matrix. The process of creation of a matrix enclosing proteins goes through the dehydration of a polymer gel that, depending on the polymers used and the amount of water retained can affect the structural stability and optical properties of the fluorescent proteins. In this contribution, we use molecular dynamics simulations to elucidate the dehydration process in a realistic model of the protein-polymer box system and its impact on the structural and optical properties of the protein. For this, we have simulated the polymer matrix using a 4:1 mix in weight of PEO and TMPE in aqueous solution. Each PEO chain has a molecular weight of 4 millions. This has been done using a novel technique to efficiently obtain a relaxed box with a desired level of crystallinity. The second part of the work involves using a novel artificial intelligence algorithm that ranks water molecules based on the likelihood of evaporation. The artificial intelligence algorithm takes full advantage of the SOAP descriptors to describe the water molecules and a K-means clustering method to classify the water molecules.

Resume : Kagome metals have recently attracted attention due to their fascinating electronic structures. Given their special geometry and depending on the degree of electron filling in their lattices, these materials are predicted to host a variety of instabilities like superconductivity (SC), spin liquid states, charge density waves (CDW), and more. One of the recently discovered non-magnetic kagome metals is AV3Sb5 (A = K, Rb, Cs). This family of layered metals exhibits CDW, unconventional superconductivity, and non-trivial band topology which believed to arise due to the proximity of saddle points in the vicinity of Fermi level. Among all the compounds, CsV3Sb5 is the heaviest member of the family which exhibits SC at 2.5 K and a CDW transition at 94 K. The charge density wave ordering manifests as a breathing-mode distortion in the kagome layers. It has been suggested that such ordering derives from nesting between saddle points on the Fermi surface. To aid in the understanding of this materials class, we present calculations of Fermi surface nesting and Lindhard susceptibility of CsV3Sb5. Additionally, we experimentally probed the coupling between the CDW and SC states via hole doping the system. The resulting phase diagram for CsV3Sb5?xSnx reveals that small carrier doping can have dramatic impacts on SC and CDW order in CsV3Sb5.

Resume : Kagome metals have recently attracted attention due to their fascinating electronic structures. Given their special geometry and depending on the degree of electron filling in their lattices, these materials are predicted to host a variety of instabilities like superconductivity (SC), spin liquid states, charge density waves (CDW), and more. One of the recently discovered non-magnetic kagome metals is AV3Sb5 (A = K, Rb, Cs). This family of layered metals exhibits CDW, unconventional superconductivity, and non-trivial band topology which believed to arise due to the proximity of saddle points in the vicinity of Fermi level. Among all the compounds, CsV3Sb5 is the heaviest member of the family which exhibits SC at 2.5 K and a CDW transition at 94 K. The charge density wave ordering manifests as a breathing-mode distortion in the kagome layers. It has been suggested that such ordering derives from nesting between saddle points on the Fermi surface. To aid in the understanding of this materials class, we present calculations of Fermi surface nesting and Lindhard susceptibility of CsV3Sb5. Additionally, we experimentally probed the coupling between the CDW and SC states via hole doping the system. The resulting phase diagram for CsV3Sb5?xSnx reveals that small carrier doping can have dramatic impacts on SC and CDW order in CsV3Sb5.

Resume : At the beginning of the century, working with first-principles molecular dynamics to characterize and understand materials was limited to systems made of 100 atoms at most, not requiring a too high number of plane waves as a basis set and for time trajectories of a few ps. Effective potentials not accounting explicitly for chemical bonding were quite often preferred in spite of their intrinsic limits. Overall, there was still a great deal of skepticism on the idea that molecular dynamics could complement experiments and even go beyond them, despite the fact that the number of users and papers had already taken a quite sharp positive slope. Due to the increase in computational power and the improved reliability of available DFT schemes, the situation has evolved dramatically in recent years, to the point of making possible the prediction of properties on systems and trajectories comparable to those employed within classical MD 20-30 years before. This talk will provide some insight into the development of first-principles molecular dynamics over the past twenty years, by taking advantage of a number of results obtained for disordered systems and nanostructures. Some methodological features will be especially highlighted, like the proper choice of the exchange-correlation functionals and the inclusion of dispersion forces.

Resume : Bi-dimensional (2D) semiconductors are central to many applications such as electronics, photocatalysis, energy harvesting. Their outstanding characteristics include high surface area, high exciton binding energy and mechanical properties enabling band gap engineering [1]. The combination of two semiconductors (heterostructure) is a good way to lift many technological deadlocks. Most 2D/2D heterostructures are van der Waals (vdW) heterostructures characterized by a weak interaction between the layers [2]. While ab initio calculations are useful to study physical properties of materials, the computational effort-accuracy tradeoff is limiting their application to few samples. Machine Learning (ML) methods are alleviating the computational demand and is applied in many fields both experimental and computational [3]. This work aims at using ML to select relevant candidates for further study through experimental work or computational calculation. First, a feature space representing the whole dataset of heterostructure samples (heterostructure space) is created. The engineered features are relating to different aspects of the crystal structure such as atomic charge and distribution, two important characteristics to assess the potential combinations of materials. Then, a meta-estimator is conceived and trained to predict feature values of samples having a defined band alignment. The chosen base model is k-nearest neighbors (KNN) regression, a robust non-generalizing model. The first degree of refinement is the separated training and convolution of four different KNN predicting different features. To further refine the regression, swarm intelligence principles [4] are used: several predictions of the same KNN convolution are collected and their boundaries are shared between each other, reducing the variance of the meta-estimator and finding regions of higher target density. The proposed algorithm is conceived with several degrees of versatility to tune the selection criteria: a tolerance factor on the value of band alignment and a sampling term to probe different parts of the heterostructure space. This new “swarm smart” algorithm is a powerful tool to select, among experimentally existing, computationally studied and not yet discovered vdW heterostructures, the most likely candidate materials enabling us to face the scientific challenges ahead. References:[1] Roldán R. et. al. (2017) Theory of 2D crystals: graphene and beyond. Chem. Soc. Rev., 46, 4387 [2] Geim A. K., Grigorieva I. V. (2013) Van der Waals heterostructures. Nature, 499, 419 [3] Strieth-Kalthoff F. (2020) Machine learning the ropes: principles, applications and directions in synthetic chemistry. Chem. Soc. Rev., 49, 6154 [4] del Valle Y. et. al. (2008) Particle Swarm Optimization: Basic Concepts, Variants and Applications in Power Systems. IEEE T. Evolut. Comput., 12, 2, 171

Resume : One of the key issues in the physics of topological insulators is whether the topologically non-trivial properties survive at finite temperatures and, if so, whether they disappear only at the temperature of topological gap closing. Here we study this problem, using quantum fidelity as a measure, by means of ab-initio methods supplemented by an effective dissipative theory built on the top of the ab-initio electron and phonon band structures. In the case of SnTe, the prototypical crystal topological insulator, we reveal the presence of a characteristic temperature, much lower than the gap-closing one, that marks a loss of coherence of the topological state. The transition is not present in a purely electronic system but it appears once we invoke coupling with a dissipative bosonic bath. Features in the dependence with temperature of the fidelity susceptibility can be related to changes in the band curvature, but signatures of a topological phase transition appear in the fidelity only though the non-adiabatic coupling with soft phonons. Our argument is valid in valley topological insulators, but in principle can be generalized to the broader class of topological insulators which host any symmetry-breaking boson.

Resume : Disordered chalcogenides have a plethora of high-technological applications, from optoelectronic devices to computing memories. In particular, the ternary Ge-Sb-Te (GST) system received widespread interest in the last 20 years due to its application in so-called Phase Change Memory (PCM) which is a promising non-volatile memory (NVM)), they have many applications start from memories to neuromorphic computing and storage applications, it exhibits a high contrast between amorphous and crystalline phases [1], this type of application leans on the difference’s behaviour of amorphous and crystalline phase, the two phases display resistivity contrast, which is exploited in phase-change memory devices. In this scope, quantum modelling has led to a significative advanced in the understanding of its disordered phase over the last decade. However, if there is somehow a certain degree of agreement nowadays by many studies on its complex amorphous structure, other challenges still require further work: one being the thermal properties behaviour of such a system at the nanoscale. In this work as the first step, we investigate the structural properties of amorphous (GeTe)x(Sb2Te3)x-1 (GST225) including structure factors, pair distribution functions, and coordination numbers, and thermal properties (thermal conductivity) and compare our results with experimental findings. The goal of this work is to discuss the system produced by different FPMD schemes [2 and suggest the best one in terms of comparison with the experimental data. In a second step, we test the performance of a Machine Learning potential (MLP) based on Gaussian Approximation Potentials (GAP) scheme [3] for GST225 already developed and available [4] and one MLP-GAP developed in-house. For the prediction of thermal properties, the FPMD and MLP schemes are combined with the so-called approach-to-equilibrium MD (AEMD) [5] to cover the size scale of interest for practical applications. References: [1] Zhang et al. Nature Rev. Mater. 4, 150 2019. [2] Bouzid et al. Phys. Rev. B 96, 224204 2017. [3] Albert Bartok-Partay et al. Quantum Chemistry, 115(16), pp. 1051-1057 2015. [4] Mocanu et al. J. Phys. Chem. B 122, 8998 (2018). [5] Duong et al. RSC Advances 11, 10747 (2021).

Resume : Over the last 30 years, first-principles simulations have revealed particularly useful to explore the physics of ferroelectrics and have gone hand in hand with the development of fully-integrated and widely-accessible software packages. Many interesting problems are however beyond the scope of first-principles investigations (limited to relatively small systems and time scales) and boosted the development of second-principles methods – i.e. construction of effective models with parameters directly fitted from first-principles. A pioneering step in that direction was the development of effective hamiltonians (Heff ) relying on the concept of local mode. Then, alternative atomistic approaches have been considered, including shell models, bond-valence model or effective atomic potentials generalizing the Heff method. Although those approaches demonstrated their usefulness, their use remains relatively limited due to the absence of integrated package implementing them. Here, we will present the recently developed MULTIBINIT open software that is distributed with the ABINIT package (www.abinit.org) and provides a semi-automatized solution for the construction and use of effective atomic potentials. The power of the method for ferroelectric oxides and related compounds will be illustrated through the the modeling of various perovskites (CaTiO3, PbTiO3, BaTiO3, BaHfO3, SrTiO3), exhibiting different kinds of phase transitions and lattice properties. Work done in collaboration with J. Bieder, M. M. Schmitt and E. Bousquet and supported by F.R.S.-FNRS Belgium (project PROMOSPAN) and the European Union’s Horizon 2020 research and innovation program under grant agreement N  766726 (TSAR).

Resume : Two dimensional organic-inorganic layered perovskites have been investigated in optoelectronic devices with remarkable performances, providing greater stability compared to 3D counterparts. Their extremely flexible organic/inorganic intercalation has provided access to materials for applications also in the fields of chirality, ferroelectrics, and spintronics. Recent studies focused on the impact of organic cations on the material’s dimensionality, bandgap, and crystallographic structure, but the effect of the organic configuration on the emission properties has not yet fully investigated. In this work, we show first the building of a database of Pb-Br 2D structures that were synthesized by using ca. 50 different organic cations. Experimental data from both synthesis conditions and resulting structure and optical properties were collected in real-time and stored on an artificial-intelligence electronic notebook. This facilitated data aggregation, normalization, and visualization in real-time. We also enriched such database with the molecular fingerprints of the organic cations for generating correlation maps. We discovered a new set of white and blue-emitting platelets. We found that molecular weight and rotatable bonds of organic cations alone do not describe the periodicity of 2D layered perovskites and their optical properties. Instead, the addition of a heteroatom in different positions of the organic tail modulates the resulting colour of the emission. Our work provides experimental guidelines to design and control the emission of 2D layered perovskite through the effective integration of experimental data with machine learning tools.

Resume : Two-dimensional (2D) Van der Waals heterostructures consist of two atomically thin sheets of varying nature which often remain connected to each other mainly through Van der Waals interactions. The vast number of combinations of 2D materials, but also often-reported property-dependencies on strain and twist-angle, make the class of 2D Van der Waals heterostructures versatile. Density functional theory (DFT) has been a popular approach to investigate Van der Waals heterostructures. However, the widely and sensibly used periodic boundary conditions (BC) impose two fundamental limitations to the treatment of this specific class of systems. Firstly, the combining of two relaxed monolayers of varying nature into a single simulation box often requires one to be squeezed and/or the other to be stretched for the purpose of fulfilling BC. Secondly, rotating one monolayer relative to the other within a fixed simulation box generally breaks BC except for a limited set of angles. Consequently, all energetic minima plausibly derived by DFT for a certain 2D Van der Waals heterostructure contain strains and twist-angles imposed by BC. Yet, by varying the shape and size of the simulation box, the twist-angle and the magnitude and direction of applied strain, different combinations of strains and twist-angles become accessible. To the best of our knowledge, no previous work involving DFT-treatment of 2D Van der Waals heterostructures deeply focused on the BC-imposed limits, but instead elaborate on the most stable system out of few intuitively derived candidates. In our work, an algorithm was written which analytically derives a complete set of BC-fulfilling heterostructure candidates below selected threshold-strain and size, starting from any two monolayers. The algorithm was presented in a case-study of PC-WS2, a type II semiconducting 2D Van der Waals heterostructure of trigonal symmetry. Also, few of the physical properties of the system were elaborated on for several of the strain and twist-angle combinations. More specifically, geometric properties, electron density distributions, planar potentials, density of states and band structures were considered. In the end, the meaningfulness of the results for the different systems were discussed. Based on our findings, we suspect that elevating the awareness of the BC-imposed limits to DFT treatment of 2D Van der Waals heterostructures could allow researchers to choose more meaningful systems to be treated in their research. Regardless, our algorithm can be considered a tool for facile design of 2D Van der Waals heterostructures. Moreover, it can be considered as a viable option for database generation of 2D Van der Waals heterostructures.

Resume : The material interface is possibly the most important part of any electronic or optoelectronic device. Another cornerstone of modern device engineering is the development of heterostructures by joining materials with different characteristics. Our work [1] combines these two fields to investigate the material engineering of 2D lateral interfaces. Specifically, we employ first-principles modelling to examine the successfully grown interface between 2D TaS2 and MoS2, materials with incommensurate primitive cells. We show that the interface can have two distinct morphologies, "dislocation" or "coherent", and our investigations identify competing physical features that drive the system towards either one. The system is in a balance between the local bond symmetry and material strain. Dislocations give rise to a breakdown of local symmetry and have an energetic cost for their formation; however, preventing dislocations incurs a cost in strain energy. Using these new insights, we apply these techniques to a variety of other 2D interfaces that share structural symmetry and predict their interface transition point from "coherent" to "dislocation". This is a powerful methodology applicable to any 2D material pair that share structural symmetry, and the qualitative understanding is of great value to any experimental or theoretical study. Overall, our insights provide a guideline for controlling the morphology of the lateral interfaces between 2D materials, which should be of high importance for controlling device functionalities. [1] F. Davis & A.V. Krasheninnikov, submitted (2022).

Resume : Recent studies in ferroelectric hafnia thin films focus on their advantages over conventional perovskite ferroelectrics which broadened the application of ferroelectric materials. However, there are shortcomings in understanding ferroelectric switching, which is crucial in the operation of these devices. In this study, the polarization switching behavior of a polycrystalline hafnium zirconium oxide (HZO) under an applied electric field is simulated using a phase-field model. The model takes into account the bulk free energy density, gradient energy density, and additional contribution from the applied electric field. To reduce the complexity and high computational cost, we neglect elastic contributions in the model and consider the effective representation of the polarization state equation by merging the contribution of strain-polarization coupling with the Landau coefficients. The Landau coefficients describing the bulk free energy density for HZO are not well-defined in literature. For the polycrystalline thin film with random orientations, a novel approach is introduced to predict effective Landau coefficients from experimentally measured switching curves by combining the phase-field model with a genetic algorithm (GA). The simulated curve generated using optimized effective Landau polynomial from GA eliminates the discrepancies in remnant polarization and coercive field observed in the previous computational studies. The domain dynamics during switching predicted by phase field simulations are consistent with available experimental findings. We validate the model by accurately simulating switching curves for HZO thin films with different ferroelectric phase fractions. Furthermore, the phase-field model is also used to investigate the effects of grain morphology and crystalline texture of the thin film on the switching dynamics.

Resume : The growth of high quality, large grain single monolayer transition metal dichalcogenides, such as WS2, through Metalorganic Chemical Vapour Deposition (MOCVD) is a complex chemical process that requires a tight control on the density of nucleation points [1][2][3] during the film growth, as well as the lateral growth from these nuclei. Critical process parameters control whether multilayer growth or single layer growth is obtained[2][4]. To control the nucleation process, it is critical to understand the thermodynamic forces driving the film formation. These individual steps are unfortunately extremely challenging to unravel experimentally[5]. In this work, we will demonstrate how first-principle calculations based on DFT, paired with classical nucleation theory provide fundamental insights in identifying the MOCVD reactor operation conditions that favour the WS2 growth . We will show how the Gibbs free energies computed for different grains [6] provide the necessary information on the factors that set the film critical nucleation size and identify the temperature, pressure and composition of the reacting gases, that enable reducing the nucleation seed density. Especially, we find that controlling the ratios of certain gas-phase components directly impacts the nucleation process to the point where growth can be completely inhibited. Finally, we will show how nucleation control through careful selection of growth operation conditions can lead to the suppression of undesired by-products through modelling as well as experimental insights. [1] C. Li, T. Kameyama, T. Takahashi, T. Kaneko, and T. Kato, Sci. Rep., vol. 9, no. 1, p. 12958, 2019 [2] B. Groven et al. Chem. Mater., vol. 30, no. 21, pp. 7648–7663, Nov. 2018, [3] X. Qiang et al., Sci. Rep., vol. 11, no. 1, p. 22285, 2021, [4] J. H. Kim et al., Appl. Surf. Sci., vol. 587, p. 152829, 2022 [5] P. A. G. O’Hare, W. N. Hubbard, G. K. Johnson, and H. E. Flotow, J. Chem. Thermodyn., vol. 16, no. 1, pp. 45–59, 1984 [6] R. A. Evarestov, A. V Bandura, V. V Porsev, and A. V Kovalenko, J. Comput. Chem., vol. 38, no. 30, pp. 2581–2593, Nov. 2017

Resume : Wafer defect map images are generated by performing electrical tests on each chip on wafer. These images demonstrate specific failure patterns occurred from semi-conductor manufacturing process. Also, these patterns can cause another defects on post process, so it is crucial for engineers to classify what kind of defect patterns this wafer is early. However, it is such difficult because large-scale data sets of wafers are produced per day. So, in an attempt to automate the classification of wafer defect maps, which are currently manual dependent, various machine learning models have been introduced. Recently, Convolutional Neural Networks have been developed in computer vision society. This method shows the high performances in computer vision problems such as classifying or recognizing images. One of the state of the art methods called Inception was presented in 2015. However, this network still cannot handle the inference time for testing because of its complexity. Also, although the introduction of several deep learning models for wafer defect map classification has been continuously attempted, it has not been successfully applied to mass production environments in the real field due to problems in classification performance and computational volume. Therefore, we propose the deep learning model integrating the inception module and the skip-connection module for wafer defect map classification. This model has a fast training and inference speed with a small number of parameters, it is highly practical when processing large amounts of real-time test data in a semiconductor manufacturing environment because of its small computational volume and high classification performance. We applied our method on the real field wafer test data, and the result showed that the proposed model takes a significant improvement on inference time by over 59% with high performance compared to the baseline model. - Index Terms - convolutional neural networks, wafer defect map, classification, inception module, skip connection module

Resume : Wafer defect map images are generated by performing electrical tests on each chip on wafer. These images demonstrate specific failure patterns occurred from semi-conductor manufacturing process. Also, these patterns can cause another defects on post process, so it is crucial for engineers to classify what kind of defect patterns this wafer is early. However, it is such difficult because large-scale data sets of wafers are produced per day. So, in an attempt to automate the classification of wafer defect maps, which are currently manual dependent, various machine learning models have been introduced. Recently, Convolutional Neural Networks have been developed in computer vision society. This method shows the high performances in computer vision problems such as classifying or recognizing images. One of the state of the art methods called Inception was presented in 2015. However, this network still cannot handle the inference time for testing because of its complexity. Also, although the introduction of several deep learning models for wafer defect map classification has been continuously attempted, it has not been successfully applied to mass production environments in the real field due to problems in classification performance and computational volume. Therefore, we propose the deep learning model integrating the inception module and the skip-connection module for wafer defect map classification. This model has a fast training and inference speed with a small number of parameters, it is highly practical when processing large amounts of real-time test data in a semiconductor manufacturing environment because of its small computational volume and high classification performance. We applied our method on the real field wafer test data, and the result showed that the proposed model takes a significant improvement on inference time by over 59% with high performance compared to the baseline model. - Index Terms - convolutional neural networks, wafer defect map, classification, inception module, skip connection module

Resume : Experimental studies and following theoretical predictions indicated that rare earth nitrides may form metastable hexagonal BN-type phases in bulk materials and solid solutions with group III nitride systems under tensile strain. In this presentation, the electronic structures of hexagonal ScN, YN, and LuN, obtained with the density functional theory calculations, are discussed with particular focus on band structures. The band gaps of bulk and monolayer rare earth nitrides are predicted to be in range from 1.12 up to 2.32 eV. The monolayer YN and LuN exhibit characteristic splittings of valence bands in the K point in the Brilouin zone, related to strong spin?orbit coupling, which are similar to those present in transition metal dichalcogenides. The 2D rare earth nitrides may be promising novel materials for applications in optoelectronics as bulk, monolayers and 2D heterostructures. This work was supported by the National Science Centre (Poland) under research Grant no. 2017/26/D/ST3/00447. Calculations were performed in Wroclaw Center for Networking and Supercomputing (Project nos. 158 and 175). [1] M.J. Winiarski and D. Kowalska, Mater. Res. Express. 6, 095910 (2019). [2] M.J. Winiarski and D. A. Kowalska, Scientific Reports 10, 16414 (2020).

Resume : Experimental studies and following theoretical predictions indicated that rare earth nitrides may form metastable hexagonal BN-type phases in bulk materials and solid solutions with group III nitride systems under tensile strain. In this presentation, the electronic structures of hexagonal ScN, YN, and LuN, obtained with the density functional theory calculations, are discussed with particular focus on band structures. The band gaps of bulk and monolayer rare earth nitrides are predicted to be in range from 1.12 up to 2.32 eV. The monolayer YN and LuN exhibit characteristic splittings of valence bands in the K point in the Brilouin zone, related to strong spin?orbit coupling, which are similar to those present in transition metal dichalcogenides. The 2D rare earth nitrides may be promising novel materials for applications in optoelectronics as bulk, monolayers and 2D heterostructures. This work was supported by the National Science Centre (Poland) under research Grant no. 2017/26/D/ST3/00447. Calculations were performed in Wroclaw Center for Networking and Supercomputing (Project nos. 158 and 175). [1] M.J. Winiarski and D. Kowalska, Mater. Res. Express. 6, 095910 (2019). [2] M.J. Winiarski and D. A. Kowalska, Scientific Reports 10, 16414 (2020).

Resume : Studying electron- and X-ray-induced electron cascades in solids is essential for various research areas at free-electron laser facilities, such as X-ray imaging, crystallography, pulse diagnostics or X-ray-induced damage. To better understand the fundamental factors defining the duration and spatial size of such cascades, we investigate the electron propagation in ten solids relevant for the applications of X-ray lasers: Au, B4C, diamond, Ni, polystyrene, Ru, Si, SiC, Si3N4, W [1]. Using classical Monte Carlo in the atomic approximation, we study the dependence of the cascade size on the incident electron or photon energy and on the target parameters. The results show that an electron-induced cascade is systematically larger than a photon-induced cascade. Moreover, in contrast with the common assumption, the maximal cascade size does not necessarily coincide with the electron range. It is found that the cascade size can be controlled by a proper choice of the photon energy for a particular material. Photon energy closely above an ionization potential can essentially split the absorbed energy between two electrons -- photo- and Auger -- reducing their initial energy and thus shrinking the cascade size. Our analysis suggests a way of tailoring the electron cascades either for applications requiring small cascades with a high density of excited electrons in them, or for large-spread cascades with lower electron densities. [1] Lipp, Milov, Medvedev, J. Synchrotron Rad. 29, 323 (2022).

Resume : Studying electron- and X-ray-induced electron cascades in solids is essential for various research areas at free-electron laser facilities, such as X-ray imaging, crystallography, pulse diagnostics or X-ray-induced damage. To better understand the fundamental factors defining the duration and spatial size of such cascades, we investigate the electron propagation in ten solids relevant for the applications of X-ray lasers: Au, B4C, diamond, Ni, polystyrene, Ru, Si, SiC, Si3N4, W [1]. Using classical Monte Carlo in the atomic approximation, we study the dependence of the cascade size on the incident electron or photon energy and on the target parameters. The results show that an electron-induced cascade is systematically larger than a photon-induced cascade. Moreover, in contrast with the common assumption, the maximal cascade size does not necessarily coincide with the electron range. It is found that the cascade size can be controlled by a proper choice of the photon energy for a particular material. Photon energy closely above an ionization potential can essentially split the absorbed energy between two electrons -- photo- and Auger -- reducing their initial energy and thus shrinking the cascade size. Our analysis suggests a way of tailoring the electron cascades either for applications requiring small cascades with a high density of excited electrons in them, or for large-spread cascades with lower electron densities. [1] Lipp, Milov, Medvedev, J. Synchrotron Rad. 29, 323 (2022).

Resume : Perovskite oxides are a versatile class of materials showing exciting features such as ferroelectricity, piezoelectricity, thermoelectricity and superconductivity, just to name a few. The versatility stem from the huge number of combinations of possible cations (A and B) of the ABO3 structure. The connectivity of the oxygen in the octahedral network plays an important role in the properties of the perovskite oxide. Tilting of the oxygen octahedra alters the bond angles, which can affect the electronic band structure. An efficient way to tune and control octahedral tilting is via external strain. Octahedral tilt engineering can therefore provide an important route for designing future electronic devices.[1] Here we investigate how strain stabilizes the octahedral tilting in SrRO3 (R=Ti,Nb) using density functional theory within the generalized gradient approximation. It is found that compressive biaxial strain in SrTiO3 stabilizes octahedral tilting around the elongated out-of-plane c-axis. Furthermore, the tilting is out-of-phase, e.g., each octahedron is tilted in opposite direction to the octahedra above and below (in the out-of-plane direction). We found that the tilt angle increases with the applied strain. SrNbO3 has been recently considered as a correlated transparent conductor both in the visible and UV regime.[2,3] In contrast to SrTiO3, the d orbitals are partially filled (4d) in SrNbO3. According to our calculations, SrNbO3 shows stabilization of octahedral tilting around the out-of-plane axis under compressive biaxial strain. However, the in-phase tilt of SrNbO3 is found to have lower energy compared to the out-of-phase tilt. As the applied strain increases, the energy minima become lower and the difference between the two tilt modes increases. This suggests that the biaxial compressive strain stabilizes the octahedral tilting, and more precisely, compressive strain enhances the energy gain of the in-phase tilt relative to the out-of-phase tilt. This leads to the stabilization of the in-phase tilt in SrNbO3 under large compressive strain. Our predictions open up new design opportunities since the out-of-phase and in-phase tilts belong to different spacegroups corresponding to different symmetries. One might therefore expect different electronic states with different properties. Recently it was shown that the out-of-phase tilt exhibits gapless linear dispersion near the Fermi level.[3] Our preliminary investigation of the in-phase tilt hints that an electronic gap is introduced near the Fermi level. This suggests that SrNbO3 might exhibit both gapless and gapped linear dispersions, depending on which tilt is present. References: [1] James M. Rondinelli et al., MRS Bulletin 37.3 (2012) [2] Yoonsang Park et al., Communications Physics 3.1 (2020) [3] Jong M. Ok et al., Science Advances 7.38 (2021)

Resume : Perovskite oxides are a versatile class of materials showing exciting features such as ferroelectricity, piezoelectricity, thermoelectricity and superconductivity, just to name a few. The versatility stem from the huge number of combinations of possible cations (A and B) of the ABO3 structure. The connectivity of the oxygen in the octahedral network plays an important role in the properties of the perovskite oxide. Tilting of the oxygen octahedra alters the bond angles, which can affect the electronic band structure. An efficient way to tune and control octahedral tilting is via external strain. Octahedral tilt engineering can therefore provide an important route for designing future electronic devices.[1] Here we investigate how strain stabilizes the octahedral tilting in SrRO3 (R=Ti,Nb) using density functional theory within the generalized gradient approximation. It is found that compressive biaxial strain in SrTiO3 stabilizes octahedral tilting around the elongated out-of-plane c-axis. Furthermore, the tilting is out-of-phase, e.g., each octahedron is tilted in opposite direction to the octahedra above and below (in the out-of-plane direction). We found that the tilt angle increases with the applied strain. SrNbO3 has been recently considered as a correlated transparent conductor both in the visible and UV regime.[2,3] In contrast to SrTiO3, the d orbitals are partially filled (4d) in SrNbO3. According to our calculations, SrNbO3 shows stabilization of octahedral tilting around the out-of-plane axis under compressive biaxial strain. However, the in-phase tilt of SrNbO3 is found to have lower energy compared to the out-of-phase tilt. As the applied strain increases, the energy minima become lower and the difference between the two tilt modes increases. This suggests that the biaxial compressive strain stabilizes the octahedral tilting, and more precisely, compressive strain enhances the energy gain of the in-phase tilt relative to the out-of-phase tilt. This leads to the stabilization of the in-phase tilt in SrNbO3 under large compressive strain. Our predictions open up new design opportunities since the out-of-phase and in-phase tilts belong to different spacegroups corresponding to different symmetries. One might therefore expect different electronic states with different properties. Recently it was shown that the out-of-phase tilt exhibits gapless linear dispersion near the Fermi level.[3] Our preliminary investigation of the in-phase tilt hints that an electronic gap is introduced near the Fermi level. This suggests that SrNbO3 might exhibit both gapless and gapped linear dispersions, depending on which tilt is present. References: [1] James M. Rondinelli et al., MRS Bulletin 37.3 (2012) [2] Yoonsang Park et al., Communications Physics 3.1 (2020) [3] Jong M. Ok et al., Science Advances 7.38 (2021)

Resume : Interfacial phenomena are one of the main drivers of the performance of various functional materials. A correct description of chemistry at the electrode/electrolyte interface or precipitation, dissolution and degradation processes of solid phases is a key factor for formulation of a solid scientific basis that is essential for design of novel energy materials and understanding of the underlying processes that determine the materials performance. With the aid of atomistic modeling we investigate the role of the aqueous solvent on the various interface phenomena in selected systems of importance for electrochemistry and nuclear waste management. In particular we discuss the impact of aqueous solvent phase on: (1) the CO2 reduction reaction pathways leading to catalysis of HCOOH on different metal surfaces, (2) the oxygen evolution reaction at the NiOOH catalyst [3] and (3) dissolution/precipitation processes in Ba/RaSO4 system [4]. In particular we discuss the advantages and limitations of the available hybrid density functional theory (DFT) – implicit solvation models computational methods, including the self consistent continuum solvation (SCCS) [5,6], VASPsol [7,8] and ESM-RISM [9] approaches. [1] Fan et al., ACS Catal. 10, 10726 (2020). [2] Ringe et al., Chem. Rev Article ASAP. (2020) [3] Eslamibidgoli et al., Electrochimica Acta, 398, 139253 (2021) [4] Bracco et al., J. Phys. Chem. C 121, 12236 (2017) [5] Andreussi et al., J. Chem. Phys. 136, 064102 (2012) [6] Giannozzi et al., J. Phys-Condens. Mat. 29, 465901 (2017) [7] Mathew et al., J. Chem. Phys. 140, 084106 (2014) [8] Mathew et al., J. Chem. Phys. 151, 234101 (2019) [9] Otani et al., Phys. Rev. B 96, 115429 (2017)

Resume : Interfacial phenomena are one of the main drivers of the performance of various functional materials. A correct description of chemistry at the electrode/electrolyte interface or precipitation, dissolution and degradation processes of solid phases is a key factor for formulation of a solid scientific basis that is essential for design of novel energy materials and understanding of the underlying processes that determine the materials performance. With the aid of atomistic modeling we investigate the role of the aqueous solvent on the various interface phenomena in selected systems of importance for electrochemistry and nuclear waste management. In particular we discuss the impact of aqueous solvent phase on: (1) the CO2 reduction reaction pathways leading to catalysis of HCOOH on different metal surfaces, (2) the oxygen evolution reaction at the NiOOH catalyst [3] and (3) dissolution/precipitation processes in Ba/RaSO4 system [4]. In particular we discuss the advantages and limitations of the available hybrid density functional theory (DFT) – implicit solvation models computational methods, including the self consistent continuum solvation (SCCS) [5,6], VASPsol [7,8] and ESM-RISM [9] approaches. [1] Fan et al., ACS Catal. 10, 10726 (2020). [2] Ringe et al., Chem. Rev Article ASAP. (2020) [3] Eslamibidgoli et al., Electrochimica Acta, 398, 139253 (2021) [4] Bracco et al., J. Phys. Chem. C 121, 12236 (2017) [5] Andreussi et al., J. Chem. Phys. 136, 064102 (2012) [6] Giannozzi et al., J. Phys-Condens. Mat. 29, 465901 (2017) [7] Mathew et al., J. Chem. Phys. 140, 084106 (2014) [8] Mathew et al., J. Chem. Phys. 151, 234101 (2019) [9] Otani et al., Phys. Rev. B 96, 115429 (2017)

Resume : The constant increase in electricity consumption forces the dynamic development of a new generation of nuclear and thermonuclear reactors hence, new materials resistant to specific operating conditions (high temperature, corrosive environment). Despite the other parameters, one of the most significant factors having the biggest impact on the materials (fuel, cladding, etc.) considered for nuclear power plants is the resistance to the creation of point and extended defects (including dislocations) coming from irradiation and fission products incorporation. The model of dislocations (based on the Peierls-Nabarro approach) assumes that the bending of atomic planes adjacent to an extra half-plane follows the arctan function. The coefficients of the function that decrease with the distance from the dislocation core are usually determined using high-resolution Transmission Electron Microscopy (TEM). However, since TEM analysis is destructive, expensive, and time-consuming, there is a need to determine dislocation parameters more reasonably. In this work, the recent studies on the modeling of edge dislocations and dislocation loops using Molecular Dynamics (LAMMPS code) for uranium dioxide that is currently used in nuclear reactors and nickel, one of the most promising materials regarding possible usage in a new generation of power plants are presented. The single-crystal structures were deformed by introducing extra half-planes of atoms and relaxed to stabilize the system. Bent planes were fitted using the arctan function to determine the function coefficients. Eventually, the parameters of dislocations were used in the McChasy 1.0 code to simulate RBS/C spectra recorded for irradiated UO2 and Ni single crystals.

Resume : The constant increase in electricity consumption forces the dynamic development of a new generation of nuclear and thermonuclear reactors hence, new materials resistant to specific operating conditions (high temperature, corrosive environment). Despite the other parameters, one of the most significant factors having the biggest impact on the materials (fuel, cladding, etc.) considered for nuclear power plants is the resistance to the creation of point and extended defects (including dislocations) coming from irradiation and fission products incorporation. The model of dislocations (based on the Peierls-Nabarro approach) assumes that the bending of atomic planes adjacent to an extra half-plane follows the arctan function. The coefficients of the function that decrease with the distance from the dislocation core are usually determined using high-resolution Transmission Electron Microscopy (TEM). However, since TEM analysis is destructive, expensive, and time-consuming, there is a need to determine dislocation parameters more reasonably. In this work, the recent studies on the modeling of edge dislocations and dislocation loops using Molecular Dynamics (LAMMPS code) for uranium dioxide that is currently used in nuclear reactors and nickel, one of the most promising materials regarding possible usage in a new generation of power plants are presented. The single-crystal structures were deformed by introducing extra half-planes of atoms and relaxed to stabilize the system. Bent planes were fitted using the arctan function to determine the function coefficients. Eventually, the parameters of dislocations were used in the McChasy 1.0 code to simulate RBS/C spectra recorded for irradiated UO2 and Ni single crystals.