Pure2DopeNet: Multimodal Neural Network-Based Predictive Modeling of Nanoparticle Properties from Pure Compounds

← Back to publications

Pure2DopeNet: Multimodal Neural Network-Based Predictive Modeling of Nanoparticle Properties from Pure Compounds

Journal Article

November 26, 2024

Can Polat, Mustafa Kurban, and Hasan Kurban – IOP’s Machine Learning: Science and Technology

Quantum chemistry is pivotal in exploring the behaviors of substances at atomic and molecular levels [1, 2]. Key simulation techniques, such as density functional theory (DFT) [3] and its more efficient variant, density functional tight binding (DFTB) [4], have become essential tools. DFT, based on quantum mechanics principles [5], provides a comprehensive framework for studying the electronic structure of materials, revealing chemical bonding, charge distribution, and other quantum phenomena. However, the computational demands of solving the many-body Schreodinger equation (SE) limit the size of systems that can be feasibly studied [6].

DFTB offers a computationally lighter alternative, allowing the study of larger and more complex systems. While it is less accurate than DFT, DFTB’s efficiency facilitates simulations of extensive molecular systems and materials [7]. However, the use of empirical parameters in DFTB, which may not be universally applicable, presents challenges in ensuring broad reliability across different materials [8].

(a): Doping process of C atoms into TiO2 compound. (b): Doping of S atoms into ZnO compound. Shared structures include doping level 0 (pure – L0), L1, L3, and L6 for both of the compounds (from top to bottom).

In addressing existing challenges in materials science, this work introduces Pure2DopeNet, a novel multimodal ML approach designed to reduce the computational requirements of material simulations and minimize the need for extensive domain knowledge. Pure2DopeNet combines text- guidance with molecular images of pure compounds to predict multiple physical parameters of doped compounds, focusing on Carbon (C) doped TiO2 [ 45] and Sulfur (S) doped ZnO [46].

The model consists of three fundamental components. A CNN based compound encoder which takes compound images and generates their embeddings. A text encoder which has two stages consists of transformer and MLP stages. Lastly, a MLP layer which is guided by the text encoder while taking the compound embeddings as input.

Key contributions of this work include: (i) the introduction of a multimodal framework integrating text and image data for compound property prediction; (ii) the ability to predict doped compound properties based on pure form simulations, extending predictive capabilities; (iii) a significant reduction in computational costs compared to traditional methods; and (iv) enhanced accessibility by minimizing the need for complex simulations, broadening its applicability in materials science.

Cite this article (BibTeX):

@article{polat2024multimodal,
	author={Polat, Can and Kurban, Mustafa and Kurban, Hasan},
	title={Multimodal Neural Network-Based Predictive Modeling of Nanoparticle Properties from Pure Compounds},
	journal={Machine Learning: Science and Technology},
	year={2024},
}

References

[1] Attila Szabo and Neil S Ostlund. Modern quantum chemistry: introduction to advanced electronic structure theory. Courier Corporation, 2012.

[2] Peter W. Atkins and Ronald S. Friedman. Molecular quantum mechanics. Oxford University Press, 2005.

[3] Walter Kohn and Lu J. Sham. Self-consistent equations including exchange and correlation effects. Physical review, 140:A1133, 1965.

[4] G Seifert. Tight-binding density functional theory: an approximate kohn- sham dft scheme. The Journal of Physical Chemistry A, 111(26):5609–5613, 2007.

[5] Jun John Sakurai and Eugene D Commins. Modern quantum mechanics, revised edition. American Association of Physics Teachers, 1995.

[6] Jie Pan. Scaling up system size in materials simulation. Nature Computational Science, 1:95–95, 2021.

[7] Mathias Rapacioli, Aude Simon, Leo Dontot, and Fernand Spiegelman. Extensions of dftb to investigate molecular complexes and clusters. physica status solidi (b), 249(2):245–258, 2012.

[8] Breno RL Galvao, Lu´ıs P Viegas, Dennis R Salahub, and Maicon P Louren¸co. Reliability of semiempirical and dftb methods for the global optimization of the structures of nanoclusters. Journal of Molecular Modeling, 26:1–8, 2020.

[46] Weilai Yu, Jinfeng Zhang, and Tianyou Peng. New insight into the enhanced photocatalytic activity of n-, c-and s-doped zno photocatalysts. Applied Catalysis B: Environmental, 181:220–227, 2016.

Pure2DopeNet: Multimodal Neural Network-Based Predictive Modeling of Nanoparticle Properties from Pure Compounds

Share this: