arXiv subjects

Vadim Korolev

Publications and source records attributed to Vadim Korolev.

8 recordsLinked to original sources

arXiv2025

Enhancing composition-based materials property prediction by cross-modal knowledge transfer

Crystal graph neural networks are widely applicable in modeling experimentally synthesized compounds and hypothetical materials with unknown synthesizability. In contrast, structure-agnostic predictive algorithms allow exploring previously inaccessible domains of chemical space. Here we present a universal approach for enhancing composition-based materials property prediction by means of cross-modal knowledge transfer. Two formulations are proposed: implicit transfer involves pretraining chemical language models on multimodal embeddings, whereas explicit transfer suggests generating crystal structures and implementing structure-aware predictors. The proposed approaches were benchmarked on LLM4Mat-Bench and MatBench tasks, achieving state-of-the-art performance in 25 out of 32 cases. In addition, we demonstrated how another modeling aspect of chemical language models - interpretability - benefits from applying a game-theoretic approach, which is able to incorporate high-order feature interactions.

cond-mat.mtrl-sci↗

arXiv2024

Data-driven prediction of structure of metal-organic frameworks

Crystal structure prediction (CSP) has proven to be a highly effective route for discovering new materials. Substantial advancements have been made in CSP of inorganic and molecular crystals, while hybrid materials, including metal-organic frameworks (MOFs), have been unfairly overlooked. The ab initio techniques adopted for the CSP of MOFs cannot be scaled to a high-throughput regime, which is required for efficient exploration of the immense chemical space. Here, we propose a data-driven method to tackle current needs of computational MOF discovery. By examining CSP through the lens of reticular chemistry, coarse-grained neural networks were implemented to predict underlying net topology of crystal graphs. The models showed satisfactory performance, which was next enhanced by limiting the applicability domain. Flue gas separation was used as an illustrative example to validate the proposed CSP approach. Several hundred in silico-generated systems revealed a notable discrepancy in adsorption capacity among competing polymorphs.

cond-mat.mtrl-sci↗

arXiv2023

Coarse-grained crystal graph neural networks for reticular materials design

Reticular materials, including metal-organic frameworks and covalent organic frameworks, combine relative ease of synthesis and an impressive range of applications in various fields, from gas storage to biomedicine. Diverse properties arise from the variation of building units$\unicode{x2013}$metal centers and organic linkers$\unicode{x2013}$in almost infinite chemical space. Such variation substantially complicates experimental design and promotes the use of computational methods. In particular, the most successful artificial intelligence algorithms for predicting properties of reticular materials are atomic-level graph neural networks, which optionally incorporate domain knowledge. Nonetheless, the data-driven inverse design involving these models suffers from incorporation of irrelevant and redundant features such as full atomistic graph and network topology. In this study, we propose a new way of representing materials, aiming to overcome the limitations of existing methods; the message passing is performed on a coarse-grained crystal graph that comprises molecular building units. To highlight the merits of our approach, we assessed predictive performance and energy efficiency of neural networks built on different materials representations, including composition-based and crystal-structure-aware models. Coarse-grained crystal graph neural networks showed decent accuracy at low computational costs, making them a valuable alternative to omnipresent atomic-level algorithms. Moreover, the presented models can be successfully integrated into an inverse materials design pipeline as estimators of the objective function. Overall, the coarse-grained crystal graph framework is aimed at challenging the prevailing atom-centric perspective on reticular materials design.

cond-mat.mtrl-sci↗

arXiv2023

Artificial bees collect diverse conformers of small organic molecules

The conformational mobility of organic molecules defined as a variability of practically accessible conformers plays a critical role in determining electronic, chemical, and physical properties within computational methods. At the same time, there is a challenge of identifying compact set of global and local conformers for the comprehensive description of potential energy surface. Here we apply nature-inspired algorithms to resolve this issue. Among all the considered algorithms, the artificial bee colony optimizer exhibits the highest performance in discovering conformers detected both in gas and condensed phases. We hope that our approach enables researchers to make a next step in organic crystal packing studies.

physics.chem-ph↗

arXiv2023

Toward Accurate Interpretable Predictions of Materials Properties within Transformer Language Models

Property prediction accuracy has long been a key parameter of machine learning in materials informatics. Accordingly, advanced models showing state-of-the-art performance turn into highly parameterized black boxes missing interpretability. Here, we present an elegant way to make their reasoning transparent. Human-readable text-based descriptions automatically generated within a suite of open-source tools are proposed as materials representation. Transformer language models pretrained on 2 million peer-reviewed articles take as input well-known terms, e.g., chemical composition, crystal symmetry, and site geometry. Our approach outperforms crystal graph networks by classifying four out of five analyzed properties if one considers all available reference data. Moreover, fine-tuned text-based models show high accuracy in the ultra-small data limit. Explanations of their internal machinery are produced using local interpretability techniques and are faithful and consistent with domain expert rationales. This language-centric framework makes accurate property predictions accessible to people without artificial-intelligence expertise.

cond-mat.mtrl-sci↗

arXiv2021

Size doesn't matter: predicting physico- or biochemical properties based on dozens of molecules

The use of machine learning in chemistry has become a common practice. At the same time, despite the success of modern machine learning methods, the lack of data limits their use. Using a transfer learning methodology can help solve this problem. This methodology assumes that a model built on a sufficient amount of data captures general features of the chemical compound structure on which it was trained and that the further reuse of these features on a dataset with a lack of data will greatly improve the quality of the new model. In this paper, we develop this approach for small organic molecules, implementing transfer learning with graph convolutional neural networks. The paper shows a significant improvement in the performance of models for target properties with a lack of data. The effects of the dataset composition on model quality and the applicability domain of the resulting models are also considered.

cs.LG↗

arXiv2020

Transferable and extensible machine learning derived atomic charges for modeling hybrid nanoporous materials

Nanoporous materials have attracted significant interest as an emerging platform for adsorption-related applications. The high-throughput computational screening became a standard technique to access the performance of thousands of candidates, but its accuracy is highly dependent on a partial charge assignment method. In this study, we propose a machine learning model that can reconcile the benefits of two main approaches-the high accuracy of density-derived electrostatic and chemical charge (DDEC) method and the scalability of charge equilibration (Qeq) method. The mean absolute deviation of predicted partial charges from the original DDEC counterparts archive an excellent level of 0.01e. The model, initially designed for metal-organic frameworks, is also capable of assigning charges to another class of nanoporous materials, covalent organic frameworks, with acceptable accuracy. Adsorption properties of carbon dioxide, calculated by means of machine learning derived charges, are consistent with the reference data obtained with DDEC charges.

physics.comp-ph↗

arXiv2019

Graph convolutional neural networks as "general-purpose" property predictors: the universality and limits of applicability

Nowadays the development of new functional materials/chemical compounds using machine learning (ML) techniques is a hot topic and includes several crucial steps, one of which is the choice of chemical structure representation. Classical approach of rigorous feature engineering in ML typically improves the performance of the predictive model, but at the same time, it narrows down the scope of applicability and decreases the physical interpretability of predicted results. In this study, we present graph convolutional neural networks (GCNN) as an architecture that allows to successfully predict the properties of compounds from diverse domains of chemical space, using a minimal set of meaningful descriptors. The applicability of GCNN models has been demonstrated by a wide range of chemical domain-specific properties. Their performance is comparable to state-of-the-art techniques; however, this architecture exempts from the need to carry out precise feature engineering.

physics.comp-ph↗