DeepModeling

Recently, the research group led by Mingxing Chen from the College of Physics and Electronic Science at Hunan Normal University investigated the evolution of flat bands in the MoSe₂/WSe₂ moiré lattices of transition metal dichalcogenides using deep learning potential and Hamiltonian models. In this study, they employed the domestic first-principles software ABACUS, which is based on the linear combination of atomic orbitals (LCAO) basis set, to prepare machine learning datasets. Subsequently, they used DeepH, based on E3 equivariant graph neural networks, to train machine learning models for predicting the moiré lattice Hamiltonian. Diagonalizing the machine learning Hamiltonian yielded the electronic structure of the moiré system. Thereafter, they utilized the self-developed band unfolding program KPROJ to project the wavefunctions of the moiré supercell onto the k-points of the primitive cell and obtained the unfolded band structure, which was used to study the influence of the twist angle on the band structure of transition metal dichalcogenides. This work revealed that the flat bands originating from the valence band edge can stem from the Γ and K valleys. When the H-stacked MoSe₂/WSe₂ moiré lattice has a twist angle of 3.89˚ and spin-orbit coupling (SOC) is not considered, a flat band with a bandwidth of approximately 5 meV appears below the valence band edge at the K point. Subsequently, as the twist angle decreases, this flat band rapidly shifts upward and becomes approximately 20 meV higher than the valence band at the K point. When the twist angle further decreases to 1.7˚, multiple flat bands emerge. The research found that spin-orbit coupling leads to a large spin splitting, comparable to that in the untwisted system (approximately 0.45 eV), and is almost independent of twisting and stacking. Therefore, in the presence of spin-orbit coupling, the flat band in the K valley remains at the top of the valence band. The band unfolding results indicate that the flat bands formed by the Γ valley and the K valley exhibit distinct responses to the twist angle. The Γ valley flat band is highly sensitive to the interlayer coupling and thus rapidly shifts upward as the twist angle decreases. Conversely, the K valley flat band has a weaker dependence on the interlayer coupling and is primarily affected by structural reconstruction. Hence, a relatively small angle (2.13˚) is required to generate the K valley flat band. As the twist angle decreases, this flat band transforms from a honeycomb lattice to a triangular lattice. The relevant research, titled "Evolution of flat bands in MoSe₂/WSe₂ moiré lattices: A study combining machine learning and band unfolding methods", was published in Physical Review B [Phys. Rev. B 110, 235410 (2024)]. Doctoral students Shengguo Yang and Jiaxin Chen are the first and second authors respectively, and Mingxing Chen is the corresponding author.

Recently, the research group led by Shi Liu from the Department of Physics, School of Science, Westlake University, based on their previous work "Modular development of deep potential" (ModDP) [1], utilized the deep potential of the solid solution Pb_x Sr_{1 - x} TiO₃ to simulate the full - composition superlattices of the (PbTiO₃)₁₀/(Pb_x Sr_{1 - x} TiO₃)₁₀ system within the range of 0 ≤ x ≤ 1. This revealed a rich phase diagram derived from topological structures. In this study, the researchers took the typical system of (PbTiO₃)₁₀/(Pb_x Sr_{1 - x} TiO₃)₁₀ as a starting point to simulate the special phase transition of vortex domains in the superlattice during the heating process, namely, the ferroelectric - like - antiferroelectric - like - paraelectric phase transition. Moreover, the ferroelectric - like phase and the antiferroelectric - like phase can be respectively regulated under an external electric field, thus achieving two different types of electric hysteresis loops. By introducing Pb doping into the layer, the researchers also found that due to the weakening of the depolarization field, a topological phase transition from the vortex state to the skyrmion state can be induced. The relevant research results, titled "Topological phase transitions in perovskite superlattices driven by temperature, electric field, and doping", were published in Physical Review B [2]. Doctoral student Jiyuan Yang is the first author.

Recently, doctoral student Yu Liu and Researcher Mohan Chen from the AI for Science Institute, Beijing and Peking University, used the domestic open-source density functional theory software ABACUS to conduct an in-depth study on the impact of multiple hyperuniform long-range order on the electronic properties of high-entropy two-dimensional material MXene. The two-dimensional materials optimized by density functional theory contain more than 1,500 atoms, indicating that ABACUS is suitable for handling large material systems. The relevant research results, titled "Multihyperuniformity in high-entropy MXenes", were published in Appl. Phys. Lett. 126, 013101 (2025) and were selected as Editor's pick by the editor (https://doi.org/10.1063/5.0246719).

DeepFlame is an open-source platform for combustion fluid calculations developed for the AI for Science era [1 - 3]. It aims to break through the long-standing application implementation challenges of traditional Computational Fluid Dynamics (CFD) in the field of combustion fluids. Since its release, DeepFlame has attracted the interest and attention of both academia and industry, and has drawn a group of outstanding developers and users, providing continuous impetus for the sustainable development of DeepFlame.

In recent years, flow and combustion problems driven by buoyancy, such as fire research in fields like batteries, have garnered increasingly extensive attention from academia and industry. The DeepFlame team has seized on this hot topic. Referring to the fireFoam solver in OpenFOAM, they constructed the dfBuoyancyFoam solver based on the PIMPLE algorithm. This solver is suitable for solving turbulent flow and diffusion flame problems driven by buoyancy, expanding the application scenarios of DeepFlame. Moreover, in the development trend of new-generation CPU supercomputers based on the ARM architecture, the newly released version has been comprehensively adapted and optimized for DeepFlame on Kunpeng and Fujitsu hardware, significantly enhancing computational efficiency and parallel scalability.

Recently, teams led by Professor Yongbing Xu from Nanjing University, Professor Lin Miao from Southeast University, and Professor Lixin He from the Hefei Comprehensive National Science Center Institute of Artificial Intelligence collaborated. By combining Mott polarization measurement technology and first-principles calculations, they found that the two-dimensional vdw magnet CrTe2 has significantly enhanced spin polarization in its ultrathin limit. Especially at a thickness of 3 layers, the spin polarization rate reaches 23.4%. The researchers used the domestic open-source density functional theory software ABACUS (Atomic Abacus) + TB2J software to calculate in detail the magnetic anisotropy energy and interlayer exchange interaction of CrTe2 thin films at different thicknesses, highlighting the relevance of perpendicular magnetic anisotropy, interlayer interaction, and electron itinerant behavior to spin polarization at different film thicknesses. The research results "Substantially Enhanced Spin Polarization in Epitaxial CrTe2 Quantum Films" were published in the Advanced materials journal.

Paper link: https://onlinelibrary.wiley.com/doi/10.1002/adma.202411137

Recently, the experimental team led by Yuanyuan Shi from the University of Science and Technology of China collaborated with the theoretical team from the Hefei Comprehensive National Science Center Institute of Artificial Intelligence. They controlled the formation of conductive filaments by using a single layer of MoS₂ as an interlayer in the TiOₓ resistive switching layer of analog resistive random access memory (analog RRAM). This improved the device's switching uniformity, linearity, and symmetry. The voltage difference between switching cycles was only 1.28% and 1.7%. This structure also achieved high conductance modulation linearity and 64 conductance states (6-bit), and the on-chip training accuracy was as high as 93.02%. The researchers used the domestic open-source density functional theory software ABACUS (Atomic Abacus) to calculate the diffusion, migration, and defect structures of Ti ions on the MoS₂ surface and successfully explained the formation mechanism of conductive filaments in RRAM based on the TiOₓ/MoS₂–xOₓ structure induced by S vacancies. This work effectively improved the uniformity, linearity, and symmetry of analog RRAM devices and provided an impetus for the development of RRAM-based in-memory computing chips. The research results "Uniformity, Linearity, and Symmetry Enhancement in TiOₓ/MoS₂–xOₓ Based Analog RRAM via S-Vacancy Confined Nanofilament" have been published in the Nano Letters journal (Nano Lett. 24, 51, 16283–16292 (2024)).

Paper link: https://pubs.acs.org/doi/10.1021/acs.nanolett.4c04434

Atomic Abacus (ABACUS) is an open-source first-principles calculation software. It supports electronic structure calculations and molecular dynamics simulations based on density functional theory. ABACUS allows users to choose plane wave basis sets or numerical atomic orbital basis sets for functions such as electronic self-consistent iterations, band structure, density of states calculations, and structural optimizations. In addition, it supports multiple DFT theoretical methods including Kohn-Sham DFT, Real-time TDDFT, Stochastic DFT, Orbital-Free DFT. ABACUS supports various exchange-correlation functionals such as local density approximation (LDA), generalized gradient approximation (GGA), metaGGA, and hybrid functionals. ABACUS also provides good support for multiple AI-assisted algorithms such as DeePKS, DeePMD, DeePTB, and DeepH. The ABACUS development team aims to support high-precision, high-stability, multi-platform, and large-scale calculations. It conducts online collaborative development in an open-source mode and ensures the rapid iteration of ABACUS in terms of functionality, performance, and reliability by constructing a high-throughput and cross-platform automated test workflow.

Virtual screening is a crucial technique in the early stage of drug discovery, aiming to identify potential drug candidate molecules from vast molecular libraries. Ligand - based virtual screening, such as molecule search, has drawn significant attention as it does not rely on specific protein site information. Recently, researchers Zhen Wang, Feng Yu, Guolin Ke, and Zhifeng Gao from DP Technology, in collaboration with Professor Zhewei Wei from Renmin University of China and doctoral student Gengmo Zhou, published a paper titled "S - MolSearch: 3D Semi - supervised Contrastive Learning for Bioactive Molecule Search" at the top machine learning conference NeurIPS 2024. This paper introduced in detail a 3D semi - supervised learning framework called S - MolSearch for active molecule search. The S - MolSearch is designed based on the principle of inverse optimal transport and can effectively combine and utilize labeled and unlabeled data, demonstrating a remarkable improvement over existing virtual screening methods. Uni - Mol functions as a 3D molecule encoder and plays a central role in this process, showcasing its great potential in molecule representation and molecule similarity measurement and providing strong support for new drug discovery.

The research group led by David Rodney from the University of Lyon 1 collaborated with Fuzhi Dai (now a distinguished professor at the University of Science and Technology Beijing) from the Beijing Institute for Scientific Intelligence. They developed a high-precision deep neural network potential function model based on the DeePMD-kit software. This model reveals the complex phase-transformation-induced plasticity (TRIP) mechanism in ceria (CeO₂)-doped zirconia (ZrO₂) ceramics. The related results were published in the authoritative journal Acta Materialia in materials science under the title "Transformation-induced plasticity in CeO₂-ZrO₂ ceramics: Atomic-scale insights using a deep neural network potential" (https://doi.org/10.1016/j.actamat.2024.120661). The model and dataset have been made public on the AIS-Square platform for researchers to use: https://www.aissquare.com/models?search=CeO₂-ZrO₂. The first author of this paper is Jinyu Zhang (now a postdoctoral fellow at Osaka University), and the corresponding author is David Rodney.

With the increasing demand for molecular representation learning in drug discovery and materials design, multimodal representation learning that combines the 3D geometric structure of molecules with biomedical text is becoming a research hotspot. Recently, a paper titled "GeomCLIP: Contrastive Geometry-Text Pre-training for Molecules" [1] published on arXiv introduced an innovative framework called GeomCLIP in detail. This framework combines the 3D geometric information of molecules with text descriptions and significantly enhances the multimodal learning ability of molecular representations through contrastive learning and denoising pre-training tasks, demonstrating superior performance in several downstream tasks.

The research team developed a large-scale dataset called PubChem3D, which contains more than 200,000 pairs of 3D geometry and text descriptions of molecules, covering rich chemical and biological information. The GeomCLIP framework adopts a dual-encoder structure to encode the geometry and text of molecules respectively and aligns the representations of the two modalities through contrastive learning while retaining the ability of the geometric encoder to model the characteristics of 3D molecules. Experiments show that GeomCLIP has achieved excellent performance in molecular property prediction, molecule-to-text retrieval, and 3D molecular caption generation tasks, providing important support for drug research and development and materials design.

This research was completed by a joint team from the Artificial Intelligence Research Laboratory at Pennsylvania State University and the Shenzhen International Graduate School of Tsinghua University. Teng Xiao from Pennsylvania State University is the first author and corresponding author of the paper, and the collaborators include Chao Cui, Huaisheng Zhu, and Professor Vasant G. Honava from Tsinghua University. This work has opened up a new direction for multimodal learning of molecular representations and also provided new solutions for drug discovery and materials science research.