Generative AI-Driven Framework for Optimal Structure Selection and Generation in Material Science

Researchers: Hamam Mokayed and Andreas Larsson

In the field of sustainable materials science, understanding and manipulating nanomaterial growth remains a daunting task. Traditional atomic scale methods such as ab initio molecular dynamics (aiMD) based on density functional theory (DFT) are often too computationally expensive to access the nano- to millisecond time scales needed to study growth mechanisms.

To overcome these limitations, we propose an AI-driven framework that simplifies the construction of machine learning force fields (MLFFs) trained on DFT results. Our recent work on developing an MLFF (DeepCNT-22), capable of modeling single-wall carbon nanotube growth, has provided new insights into the formation of interface defects and the conditions required for defect-free growth.

Aiming for further enhancement, our research seeks to expand DeepCNT-22 to incorporate hydrogen, targeting a detailed understanding of hydrocarbon feedstock gas decomposition in material growth processes. The long-term goal is to create an atomic-level digital twin of CVD/ALD material growth.

However, incorporating hydrogen introduces the challenge of efficiently selecting and generating representative atomic configurations. Our novel approach is to develop an AI-driven framework that leverages advanced machine learning models and physics-informed learning to optimize atomic environment descriptors, and to use this framework to select and generate representative configurations for MLFF training.

This development not only promises significant efficiency gains in MLFF construction, reducing reliance on computationally expensive DFT calculations, but also lays the groundwork for more extensive future explorations in materials science. A strong synergy between WISE and WASP is central to this effort.

Sustainability aspects

This framework addresses key challenges faced by MLFFs in adapting to varied chemical environments. Many existing models struggle to generalize across different systems without retraining, which is time-intensive and resource-consuming. By focusing on robust and high-quality dataset generation, our framework enhances both performance and resilience of MLFF models.

In collaboration with WISE and WASP, this work enables advances in controlling nanomaterial growth, supporting applications in energy-efficient carbon nanomaterials for electronics and energy storage. The proposed framework plays a crucial role in enabling defect-free carbon nanomaterials with consistent properties, advancing their use in sustainable technology solutions.

The scalability and efficiency of these MLFF models will set new benchmarks for computational materials research, opening pathways to address global sustainability challenges and fostering innovation in clean energy and green technologies.

Project page on the WISE website

• An AI-Driven Framework for Generating Optimal Structures in Material Science Using Generative Methods