A team under leadership of Kostya Novoselov and Andrey Ustyuzhanin published a paper “Sparse representation for machine learning the properties of defects in 2D materials” in a prestigious journal npj Computational Materials. The code and the trained model weights are available at Constructor Research Platform.
2D materials offer exciting opportunities as building blocks for new electronic devices, such as bendable screens, efficient solar panels, and high-resolution cameras. One of the defining properties of 2D materials is the high influence of crystal lattice imperfections, or defects. They radically change the electronic properties. They can turn isolators into semiconductors, semiconductors to metals, make materials magnetic or catalytic.
Ideal crystalline materials consist of an infinite repeating pattern. Real crystalline materials have defects. The work studies point defects: vacancies, when an atom is removed, and substitutions, when an atom is replaced with a different one. Example MoS2 structure with one S vacancy, one Mo vacancy and two S to Se substitutions:
In order to do in-silico engineering of defects, rapid estimation of the material properties is required. The authors developed a machine learning approach that is at least 1000 faster compared to first principles calculations and 3.7 times more accurate than competing machine learning approaches. Its main idea is using a sparse representation graph consisting only of the point defects as an input to a machine learning algorithm, as opposed to the traditional approach when all the atoms are used.
It is depicted below:
a) Start with a full structure of a 2D material with defects
b) Get the sparse structure by removing the sites that don’t contain defects
c) A graph built by connecting the defect sites that are closer than the cutoff radius
d) Resulting sparse graph. Note the edges going through the periodic boundary.
Computational research projects often evolve into software engineering projects, and this one was no exception. The core idea of the paper was invented and tested in about a week. All the remaining 1.5 years went into writing the code and running meticulous experiments.
The project was a collaboration between 4 different institutions, each using their own computational resources, which complicated the matters further. Constructor research platform arrived at the late stage of the project, and was used for:
- Running final evaluation experiments
- Training the final models for inference
- Dataset publication
- Publishing prediction interface
Value of Constructor Research Platform:
The Constructor Research platform offers immense value to researchers studying the properties of defects in 2D materials. By leveraging the platform's capabilities, researchers can apply sparse representation for machine learning, enabling rapid estimation of material properties based on lattice structure and defect configuration. Built-in versioning control ensures that research projects are organized and tracked effectively, enabling researchers to manage and compare different iterations of their experiments effortlessly. This streamlined workflow enhances productivity and facilitates the reproducibility of results.
Designed as a web-based collaborative environment, the platform offers easy onboarding for research team members. Researchers can seamlessly collaborate, share insights, and collectively contribute to the understanding of defect-engineered materials. The platform's intuitive interface and collaborative features foster a productive and inclusive research environment.
The Constructor Research platform includes project sharing capabilities, allowing researchers to share their work through interactive publications. With interactive publications, researchers can present their findings in an engaging and interactive manner, enhancing the accessibility and impact of their research.
Additionally, the platform provides shareable access to reproducible experiments, facilitating collaboration with both internal and external contributors. Researchers can easily share their experiments with collaborators, granting them access to reproduce and build upon the research findings. This collaborative approach enhances knowledge exchange, encourages interdisciplinary collaboration, and accelerates research progress.
In summary, the Constructor Research platform offers a comprehensive and collaborative environment for researchers studying the properties of defects in 2D materials. With tailored compute infrastructure, a complete set of research and development tools, web-based collaboration, project sharing capabilities, and shareable access to reproducible experiments, the platform empowers researchers to advance their investigations, foster collaboration, and drive scientific advancements in the field. Constructor Research platform proves to be highly resource-efficient, significantly reducing computational costs for both training and inference.