Samsung Science & Technology Center

Discovery of new molecules and materials with desired properties is a practical goal of chemical research. A promising way to significantly accelerate the latter process is to incorporate all available knowledge and data to plan the synthesis of the next materials. In this talk, I will present several directions to use informatics and machine learning to efficiently explore chemical space[1]. I will first describe methods of machine learning for fast and reliable predictions of materials and molecular properties[2-3]. With these tools in place for property evaluation, I will then present a few generative frameworks that we have recently developed to allow the inverse design of molecules and materials with optimal target properties, either in the compositional space or structural space[4-5]. One general challenge in digital discovery is that many of the molecules and materials that are computationally designed are often discarded in the laboratories since they are not synthesizable. I will thus spend some time to talk about the synthesizability of molecules and materials, either by predicting the synthesis pathways (retrosynthesis) or chemical reactivity[6-9]. All of these recent works, however, are only the beginning of the field with many exciting further developments to be seen. Several challenges and opportunities that lie ahead for further developments of accelerated chemical platform will finally be discussed.

In this talk, we will explore two different paths of designing and discovering novel porous materials using both natural (i.e. computational) and artificial designs. Regarding the natural design, novel composite metal-organic frameworks (MOFs) were designed using our lattice-matching algorithm and our predictions were realized via successful synthesis of various MOF@MOF core shell materials [1]. These materials possess potential synergetic properties for various applications including chemiresistor sensors [2]. Regarding the artificial design, we will talk about developments made in using the GAN model to create porous materials [3] and some of the other implementations to take advantage of the modular nature of the MOFs [4] and to build towards universal transfer learning to predict various properties of the MOFs [5]. Finally, we will provide some opinions on which of the following two paths will be the winning strategy when it comes to meaningfully extending the materials space for the porous materials. In general, the bottleneck for materials discovery in porous materials seem to come from successful experimental synthesis of the discovered materials, which can be a difficult hurdle for computational researchers and can be the main topic of discussion for the meeting.

The machine learning potential (MLP) derives total energies using machine learning models trained on reference DFT datasets. The accuracy and usefulness of MLPs are demonstrated through numerous examples of large-scale and long-term simulations across diverse application domains, such as catalysts, batteries, and semiconductors. The data-centric approach in MLPs represents a radical departure from traditional DFT calculations or classical potentials, leading to a paradigm shift in atomistic simulations and paving the way for new technological advancements. This presentation aims to provide a critical overview of the current status of MLPs and explore future research directions that have emerged due to their advancements. I will first discuss the strengths and weaknesses of MLPs, highlighting the limitations of current models. In particular, I will present the challenges in constructing training sets efficiently and effectively. I will also touch on recent notable developments in MLPs, particularly graph-based equivariant models. The computational speed of MLPs enables routine simulations for ~10 nm and ~1 ns in length and temporal scales, respectively, which is significantly more realistic than conventional DFT calculations but still way smaller or shorter than actual experiments. This raises the question, "What can we achieve with MLP-powered atomistic simulations?" I will propose some potential research directions that may have a real impact on materials engineering.

As size of electronic devices are miniaturized to nanoscale, the precision of their fabrication processes is becoming extremely demanding. Atomic layer deposition (ALD) is a vapor phase thin film deposition technique based on sequential, selflimiting surface reactions. Through ALD, exceptional conformality on high-aspect ratio structures, thickness control at the Angstrom level, and tunable film composition are achievable. Furthermore, area-selective ALD (AS-ALD) has recently emerged as a possible alternative bottom-up approach for nanoscale patterning. With these advantages, ALD is gaining interest as a powerful tool for many industrial and research applications, especially in microelectronic fabrication. Furthermore, atomic layer etching (ALE), which is conceptionally the reverse process of ALD, is emerging as a novel technique that can provide atomically controlled etching of materials.

Ideally, the entire ALD process should be based only on the surfaces of the substrates. Thus, it is important to understand the surface chemical reaction mechanisms during ALD, in order to improve the process conditions and material quality, and even to design novel materials and processes. With development of modern simulation tools, utilization of quantum mechanical calculation is becoming increasingly useful toward deeper understanding and design of such chemical reactions. However, ALD processes often face limitations toward fabrication of next-generation semiconductor devices due to their size scale and structural complexity; furthermore, such problems are often convoluted with challenges toward realistic simulations of surface chemical processes.

In this talk, analysis of fundamental surface chemistry of ALD processes based on firstprinciples calculations, as well as development of new processes and materials based on chemical simulations will be presented. Challenges in current approaches toward simulations of surface chemistry will be also discussed, which may be resolved via introduction of novel techniques at various scales.

Predicting properties of battery slurry is extremely hard due to its invisibility of nano-micro structure of ingredients having different length scales. In this study, the crystallization behavior of polyvinylidene fluoride (PVDF) in NMP/DMF solvent at 9 to 67 weight percent (wt%) was analyzed by molecular dynamics (MD) simulation. The PVDF phase did not gradually change with the incremental increase in PVDF wt%, but displayed rapid shifts at 34 and 50 wt% in both solvents. The solvation behavior between the two solvents was quite identical from the similar radial distribution functions. However, PVDFs in DMF solvent showed a higher ratio of β phase crystalline structures than those in NMP solvent. It was found that DMF solvents were more tightly packed near trans state PVDF fluorine compared to NMP solvents. Also, NMP oxygen atoms interacted more favorably with gauche state PVDF hydrogen atoms over DMF oxygen atoms. The evaluation of properties observed in atomic scale interactions, such as trans state inhibition and gauche state preference, can be used as indicators in future solvent research. At the end of the talk, water-CMC battery slurry system with MARTINI force field will be discussed.

Soft matter refers to a wide range of systems, including polymers, colloids, biological systems, and surfaces/interfaces. Due to their "soft" properties, these systems possess highly disordered characteristics, making them applicable to both scientific and industrial fields. For example, polymers are widely used in various industrial applications due to their ability to control their properties by changing their length and monomer units. These applications include not only packaging and containers but also battery separators, polymer solid electrolytes, adhesives, and drug delivery systems. In this presentation, we will discuss computer simulations of soft matter, particularly molecular dynamics simulations for studying structure-property relationships. We will also introduce the limitations and potential solutions of molecular dynamics simulations, especially in systems where polarization is important or where chemical reactions occur. Finally, we will present recent trends in applying machine learning to molecular dynamics simulations.

According to Cohen's reciprocity, processing determines the microstructure, which in turn dictates the properties of a material. Following the same logical structure, the spatial configuration of atoms (crystal structure) and the resulting electronic structure determine the physical and chemical properties of a perfect material without any defects. Additionally, the spatial arrangement of the crystal structure and defects influences the microstructure, which subsequently determines the physical and chemical properties of materials on a macro scale. Based on these hierarchical relationships, efforts to predict and design the macro-properties and manufacturing processes of new materials at the atomic scale surged globally in the early 2000s, a movement referred to as Integrated Computational Materials Engineering (ICME). However, due to a lack of computing power, interfacing technology between scales, and insufficient meso-scale microstructure prediction models, interest in ICME waned in the 2010s. Recently, with the rapid advancements in computing power and artificial intelligence technology, as well as significant progress in meso-scale microstructure prediction models, ICME has regained attention as a pre-design model for new materials and scale-up processes. This presentation will discuss the challenges associated with existing ICME approaches, how these problems have been addressed, and how ICME will be employed in the future for material and process design.

Recently, computational materials science has been regarded as an essential component to accelerate material development. Previously, computational materials science was mainly used for analytical purposes to understand experimental phenomena, which indicates that computational materials science can be classified as an analysis technology of materials. However, as a computer power greatly improves, computational materials science enables to design materials, which is in contrast to an analysis technology of materials. Despite the significant improvement in technical performance of computational materials science, there still exist gaps with experimentation in many research areas. Representatively, there are cases in which material properties predicted by computational materials science are not observed in experiments, and there are many cases where materials predicted by computational materials science cannot be synthesized in actual experiments. What efforts should we make to reduce this gap between computation and experiment? Computational materials science technologies at atomic level (e.g., first principles, molecular dynamics, etc.), which are currently widely used for material design, is a method of predicting physical properties by inputting structural (composition) information of a material. In other words, the focus is on understanding the correlation between structure and physical properties of materials. However, in order to reduce the current gap between computation and experiment, it is necessary to understand the correlation between synthesis/process and materials properties. Current computational materials science has many difficulties in dealing with the ‘synthesis/ process’ conditions of materials. An alternative to overcome this is to combine it with data science. Using data and artificial intelligence (AI), it is possible to conduct research to understand the correlation between the process-structure-physical properties of materials. As an example, in this talk, I will introduce an AI robot-based autonomous lab for nanoparticle synthesis research. And I will discuss the future direction of the autonomous lab.

Computational modeling has emerged as a powerful tool for understanding catalytic reactions at the molecular level and improving the efficiency of catalytic systems. However, the accuracy and reliability of computational models for catalysis applications still require further improvements. The discrepancy between computation and experiment often comes from computational methodology’s precision, but mostly comes from the way in which the experimental conditions are simulated in the model. Although idealized computational models are effective at predicting the intrinsic properties of materials, such as their composition and elemental combination, their predictive power decreases significantly when they do not adequately account for external factors present in experiments, such as solvent and morphology. Here, I will discuss our combinatorial studies that combine DFT calculations with experiments on metal oxide catalysts as an illustration of how computational modeling’s the predictability can be improved. These include CO oxidation or water gas shift reaction on various rare-earth metal [1,2] or transition metal [3,4,10] doped metal oxide catalysts. In addition, a noble method of co-doping is suggested to overcome the limitation of single metal doped metal oxides [4-9, 11,12]. Our results will provide useful insight to enhance and predict the catalytic activity of metal oxide-based catalysts.

Extensive research has been dedicated to the development of high-energy-density and safe all-solid-state batteries for electric vehicles, requiring the discovery of novel solid electrolytes with high ionic conductivity and advanced cathode materials with increased energy density. However, solely relying on experiments for material development and optimization in new battery systems presents significant challenges. First-principles-based atomistic modeling is a powerful tool for predicting key properties in less-explored or less-understood material systems, thereby overcoming these obstacles. Nonetheless, traditional first-principles calculations often struggle to predict specific properties accurately. For instance, the results on oxygen-redox stability in high-energy-density cathode materials depend on the density functional theory (DFT) functionals. Likewise, room-temperature ionic conductivity extrapolated from ab-initio molecular dynamics (AIMD) calculations conducted at elevated temperatures (around 1000K) may lack accuracy due to non-linear temperature dependence at lower temperatures. To overcome these limitations, researchers are exploring novel DFT methods and DFT-based machine learning potentials (MLPs), respectively. In this talk, I will share the challenges associated with conventional DFT methodologies for developing all-solid-state battery materials and discuss potential solutions to surmount these obstacles.