Open-source density functional theory software ABACUS: promoting the development of large atomic-scale machine learning pre-trained models

Abstract

Machine-learning-enriched interatomic potential frameworks are notable catalysts that transform the realm of computational materials science. By skillfully combining quantum mechanical accuracy with the computational efficiency of classical empirical potentials, these methods significantly enhance the capabilities of molecular dynamics simulations and related investigative fields. The integration of these frameworks provides researchers with the ability to analyze the complex dynamics of material systems with an unprecedented combination of accuracy and speed. The increasing application of machine learning potential function methods and the accumulating data lay a solid foundation for the construction of a large atomic model (LAM) that comprehensively covers the elements of the periodic table, turning the vision into a feasible reality. Within this paradigm, the open large atomic model (OpenLAM) project aims to build a high-precision, high-efficiency pre-trained model suitable for complex material systems. The efficient and accurate interatomic potential models can archive through further downstream fine-tuning and distillation. A key factor in this endeavor is the comprehensive collection of advanced training data, which relies on the accuracy provided by first-principles software. ABACUS, a leading open-source tool in the density functional theory landscape, plays a pivotal role in advancing the OpenLAM project. With support for plane waves and numerical atomic orbital basis sets, ABACUS enables researchers to select the most suitable basis set for their specific research needs. Additionally, ABACUS provides an accessible pseudopotential and numerical atomic orbital library, and enables users to conveniently select pseudopotentials and orbits. ABACUS also fosters a collaborative framework for first-principles computation, creating a community-oriented research environment. And at present, ABACUS has supported lots of algorithms that combine DFT and AI, such as DeepKS, DeepH, etc. ABACUS’s hardware versatility, demonstrated by its successful adaptation across various platforms, significantly enhances its utility by leveraging the computational power of diverse architectures. The successful adaptation on domestic deep computing unit (DCU) hardware makes ABACUS be a cost-effective choice. This flexibility has been instrumental in enabling ABACUS to contribute a substantial amount of density functional theory data to the OpenLAM project, covering a wide range of material systems including alloys, semiconductors, and perovskites. The use of ABACUS for generating large amounts of data in these fields has validated the stability and reliability of ABACUS, with the potential to unlock novel functionalities and applications in modern technology. The evolution of ABACUS is closely aligned with the rapidly advancing field of AI for Science. As artificial intelligence continues to integrate into scientific research, ABACUS is well-positioned to incorporate emerging AI-driven methodologies and algorithms. Moving forward, ABACUS will remain a cornerstone of the OpenLAM project and a catalyst for new discoveries, driving materials science into a new era of AI-enhanced research and development.