Reverse engineering of materials using artificial intelligence: from foundation models to functionalities

• Grant number: 24/22392-2
• Support Opportunities: Scholarships in Brazil - Doctorate (Direct)
• Start date: June 01, 2025
• End date: May 31, 2029
• Field of knowledge: Physical Sciences and Mathematics - Physics - Condensed Matter Physics
• Principal Investigator: Gabriel Ravanhani Schleder
• Grantee: Pedro Henrique Machado Zanineli
• Host Institution: Centro Nacional de Pesquisa em Energia e Materiais (CNPEM). Ministério da Ciência, Tecnologia e Inovação (Brasil). Campinas , SP, Brazil

Abstract

Technological advancement in several production sectors today depends on materials science to discover and propose new, innovative structures with specific and optimized properties.This means that the paradigm currently explored is based on reverse engineering of the process, so that the calculation of such desired characteristics enables the development of new materials within the large sample space formed by the combination of chemical elements and structural arrangements.To this end, with the constant advancement of computational power and the vast availability of data, the use of machine learning techniques is beneficial for an efficient exploration of the space of possible structures.In particular, foundation models stand out through diffusion models in the generation of stable, unique and innovative structures by transferring knowledge from large, previously trained models. However, they are still limited in the process of generalizing problems due to the properties covered.In this way, this project aims to go a step beyond the current state of the art in foundational models for materials physics, by exploring the integration of fundamental characteristics of matter for different challenges, especially the potential energy surface and electronic structure/properties. By integrating the knowledge learned through existing models, such as Machine Learning Interatomic Potentials (MLIPs) and Equivariant Graph Neural Networks, we will develop state-of-the-art foundational models that seek generalization to a range of applications through a multimodal approach, that is, the combination of different input/output modalities, which potentially lead to synergistic gains in the learning performance of the models and their corresponding predictions. By obtaining such models, we will be able not only to predict structural, dynamic, and energetic properties of different materials and systems, but also to discover and design materials with functionalities intended for applications of interest, such as in the areas of energy, catalysis, and spintronics. (AU)