DeepModeling

On May 6, 2025, the DeepModeling Community released Community Manifesto 2.0, planning to rapidly expand exploratory work in the field of "AI literature reading" in the near future. Today, the SciAssess project has officially joined the DeepModeling Community. Developed jointly by DeepSeek and the Beijing Academy of Scientific Intelligence, this system is a testing benchmark specifically designed to evaluate the scientific literature analysis capabilities of large language models (LLMs), aiming to advance the process of AI empowering scientific research. SciAssess will collaborate with the community to launch explorations in the field of AI for literature analysis.

Water, one of the most common yet complex molecules, has long perplexed researchers with its thermodynamic properties. A controversial hypothesis suggests that under supercooled conditions, water may exhibit a "liquid-liquid transition (LLT)"—a transformation between low-density liquid (LDL) and high-density liquid (HDL)—governed by a "second critical point (LLCP)."

However, direct observation of the LLCP has been extremely challenging. Within experimentally accessible temperature-pressure ranges, liquid water readily freezes into ice, while simulations struggle to reach the microsecond timescale. As a result, this hypothesis has long remained an unresolved "mystery".

Recently, F. Sciortino et al. leveraged the DeePMD framework with the DNN@MB-pol potential model to conduct microsecond-scale molecular dynamics simulations, achieving high precision approaching CCSD(T) calculations. For the first time, they provided strong constraints on the location of the liquid-liquid phase transition critical point in water. Published in Nature Physics under the title "Constraints on the location of the liquid-liquid critical point in water," this work opens a new chapter in understanding water's anomalous behavior.

Recently, the Journal of Chemical Theory and Computation published a research work titled DeePMD - kit v3: A Multiple - Backend Framework for Machine Learning Potentials [1]. This work focuses on a core innovation in the DeePMD - kit v3 version - the multi - backend framework. The latest version has integrated four deep learning backends: TensorFlow, PyTorch, JAX, and PaddlePaddle.

Recently, the Journal of Chemical Information and Modeling published a research work titled "DeePMD-GNN: A DeePMD-kit Plugin for External Graph Neural Network Potentials" [1]. DeePMD-GNN is a plugin for DeePMD-kit [2], which for the first time achieves seamless integration of external Graph Neural Network (GNN) potential energy models within the DeePMD-kit framework, including mainstream models such as NequIP and MACE.

Large atomic models (LAMs) have undergone remarkable progress recently, emerging as universal or fundamental representations of the potential energy surface defined by the first-principles calculations of atomic systems. However, our understanding of the extent to which these models achieve true universality, as well as their comparative performance across different models, remains limited. This gap is largely due to the lack of comprehensive benchmarks capable of evaluating the effectiveness of LAMs as approximations to the universal potential energy surface. 

Conformation, which refers to the different atomic arrangements a molecule can adopt in three-dimensional space, is one of the core factors determining a molecule's physicochemical properties. However, in current mainstream molecular machine learning modeling practices, conformational information is often ignored or simplified. Most prediction tasks still rely on two-dimensional structure representations, and there is even inconsistency in using the most stable conformational form. The lack of conformational information has become a significant factor limiting model prediction accuracy and also exposes the shortcomings of existing molecular representation methods in handling conformation-sensitive tasks.

Recently, the research group led by Researcher Xu Shenzhen from the School of Materials Science and Engineering at Peking University collaborated with the Beijing Academy of Intelligent Sciences (AISI) and Deep Potential Technology (DP Technology). They employed the deep potential method[1] to study the nuclear quantum effects in proton-coupled electron transfer during electrochemical reactions and develop computational methods. Notably, all first-principles calculations in this study were carried out using the domestic first-principles software ABACUS [2]. The relevant research findings were published in Nature Communications under the title "Probing Nuclear Quantum Effects in Electrocatalysis via a Machine-Learning Enhanced Grand Canonical Constant Potential Approach" [3]. Doctoral students Sun Menglin, Jin Bin, and Yang Xiaolong are the co-first authors, and Researcher Xu Shenzhen is the corresponding author.

Recently, Associate Professor Hu Zhixin from Tianjin University collaborated with the research groups of Professor Ji Wei and Associate Researcher Wang Cong from Renmin University of China. Based on first-principles calculations and using both VASP and the domestic first-principles software ABACUS, they revealed the microscopic mechanisms of the changes in the easy magnetization axis and topological properties with the number of layers in MnSe₂. The relevant research results were published in the journal Physical Review B under the title "Interlayer coupling driven rotation of the magnetic easy axis in MnSe₂ monolayers and bilayers" (DOI: 10.1103/PhysRevB.111.054422). The first authors of the paper are Zhang Zhongqin and Wang Cong.

Abstract

Recently, Zhao Haotian, a PhD candidate at the University of Science and Technology of China, and Professor He Lixin proposed and implemented a Hybrid Gauge Real-Time Time-Dependent Density Functional Theory (Hybrid gauge rt-TDDFT) applicable to atomic orbital basis sets in the domestic open-source density functional theory software ABACUS. Based on the traditional velocity gauge, this method introduces a time-varying phase dependent on the vector potential, effectively overcoming the systematic errors caused by the incompleteness of the local basis set and providing consistent and reliable simulation results in both periodic and non-periodic systems. At the same time, this method significantly improves the calculation efficiency in periodic systems, offering a new solution that combines accuracy and efficiency for first-principles real-time dynamics simulations under the action of an electric field.

Enzymes are the core catalysts of life activities, and the prediction of their functions has important applications in various fields. However, traditional methods of enzyme classification, such as EC numbers or protein family classification, rely on manual classification, which can lead to inaccurate classification granularity. Moreover, these methods lack the ability to characterize the dynamic structural transformation of substrates and products during the reaction process, making it difficult to accurately understand the actual functions and catalytic mechanisms of enzymes. Against this backdrop, ReactZyme, developed by Hua et al., brings new hope to the study of enzyme functions. Uni - Mol, as the core engine for molecular modeling, provides powerful support for capturing the complex three - dimensional interactions between enzymes and substrates. This achievement was presented at the Datasets and Benchmarks track of the 38th Conference on Neural Information Processing Systems (NeurIPS 2024), and its code and data have been open - sourced(https://github.com/WillHua127/ReactZyme).