Learning local equivariant representations for quantum operators

Authors: YinZhangHao Zhou, Zixi Gan, Shishir Pandey, Linfeng Zhang, QIANGQIANG GU

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate SLEM s capabilities across diverse 2D and 3D materials, achieving high accuracy even with limited training data. SLEM s design facilitates efficient parallelization, potentially extending DFT simulations to systems with device-level sizes, opening new possibilities for large-scale quantum simulations and high-throughput materials discovery. ... Table 1 presents a comparison of mean absolute error (MAE) values in Hamiltonian prediction for graphene, Mo S2, and Si systems...
Researcher Affiliation Collaboration Zhanghao Zhouyin2,3 Zixi Gan2,4 Shishir K. Pandey5 Linfeng Zhang2,6 Qiangqiang Gu1,2,7 1School of Artificial Intelligence and Data Science, University of Science and Technology of China, Hefei, China 2AI for Science Institute, Beijing, China 3Department of Physics, Mc Gill University, Montreal, Canada 4Department of Chemistry, Zhejiang University, Hangzhou, China 5Birla Institute of Technology & Science, Pilani-Dubai Campus, Dubai, UAE 6DP Technology, Beijing, China 7Suzhou Institute for Advanced Research, University of Science and Technology of China, Suzhou, China
Pseudocode No The paper describes the model architecture and updating rules using mathematical formulations and textual descriptions, along with a schematic diagram in Figure 2, but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes The implementation of the SLEM model and its semi-local variant (named LEM) are open-source and accessible via the Dee PTB Git Hub repository. To facilitate the integration of DFT outputs with machine learning models, we have developed and released a supplementary tool, dftio.
Open Datasets Yes The dataset used in this work is uploaded in the open-source platform AISquare via this link: https://www.aissquare.com/datasets/detail?pageType=datasets&name=Quantum_Operator_Dataset&id=286.
Dataset Splits Yes Among them, 150 frames are randomly split as the training set, and 30 frames are used for testing. ... The dataset is split as 104,001/17,495/9,335.
Hardware Specification Yes In practice, for Hf O2 with 4s2p2d2f1g basis, a typical model of 1.7M parameters can predict the quantum operators for up to 103 atoms on devices with 32GB memory.
Software Dependencies No The paper mentions several software tools used: Dee PMD, Lammps, ABACUS, and Py Torch. However, it does not provide specific version numbers for these software dependencies, which are required for full reproducibility.
Experiment Setup Yes A typical input file of the lammps sampling looks like: variable NSTEPS equal 500000 variable THERMO_FREQ equal 100 variable DUMP_FREQ equal 1000 variable TEMP equal 300 variable PRES equal 1.00 variable TAU_T equal 0.10 variable TAU_P equal 0.50 ... timestep 0.001. A typical DFT calculation input looks like: INPUT_PARAMETERS ntype 1 ecutwfc 100 scf_nmax 100 smearing_method gaussian smearing_sigma 0.002 basis_type lcao ks_solver genelpa mixing_type pulay mixing_beta 0.7 scf_thr 1e-07 out_chg 1 symmetry 1 calculation scf out_band 1 force_thr 0.001 out_stru 1 kspacing 0.08 lspinorb 0 out_wfc_lcao 0 dft_functional pbe out_mat_hs2 True