Learning Equivariant Non-Local Electron Density Functionals
Authors: Nicholas Gao, Eike Eberhard, Stephan Günnemann
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In our empirical evaluation, we find EG-XC to accurately reconstruct gold-standard CCSD(T) energies on MD17. On out-of-distribution conformations of 3BPA, EG-XC reduces the relative MAE by 35 % to 50 %. Remarkably, EG-XC excels in data efficiency and molecular size extrapolation on QM9, matching force fields trained on 5 times more and larger molecules. On identical training sets, EG-XC yields on average 51 % lower MAEs. |
| Researcher Affiliation | Academia | Nicholas Gao , Eike Eberhard , Stephan Günnemann EMAIL Department of Computer Science & Munich Data Science Institute Technical University of Munich |
| Pseudocode | No | The paper describes methods like the self-consistent field (SCF) method and equivariant message passing through textual descriptions and mathematical equations. It does not contain any clearly labeled 'Pseudocode' or 'Algorithm' blocks, nor does it present structured steps formatted like code. |
| Open Source Code | Yes | We provide the source code on https://github.com/eseberhard/eg-ex |
| Open Datasets | Yes | We compare these methods on the revised MD17 dataset, which contains precise gold-standard CCSD(T) (CCSD for aspirin) reference energies for conformations of five molecules along the trajectory of a molecular dynamic (MD) simulation (Chmiela et al., 2018). ... To investigate the extrapolation to unseen structures, we use the 3BPA dataset (Kovács et al., 2021). ... Here, we simulate this setting by splitting the QM9 dataset (Ramakrishnan et al., 2014) into subsets of increasing size based on the number of heavy atoms |
| Dataset Splits | Yes | Each molecule has a training set of 1000 structures, which we split into 950 training and 50 validation structures. Each test set contains an additional 500 structures (1000 for ethanol). ... The training set consists of 500 structures sampled from an MD simulation at room temperature (300K). The test sets consist of MD trajectories at 300K, 600K, and 1200K ... For each training set, we split the structures 90%/10% into training and validation sets. |
| Hardware Specification | Yes | All calculations were performed on a single NVIDIA A100 GPU with our JAX implementation. |
| Software Dependencies | No | The paper mentions using 'JAX (Bradbury et al., 2018)' for the SCF method and 'Py SCF (Sun et al., 2018)' for precomputing integrals and obtaining grid points. It also mentions using the Si LU activation function (Hendrycks & Gimpel, 2023). However, specific version numbers for JAX, Py SCF, or any other software libraries are not provided, only citations to their respective papers. |
| Experiment Setup | Yes | Table 4: Hyperparameters for EG-XC. d number of features per irrep 32, lmax number of irreps 2, T number of layers 3, Radial filters 32, ϵm GGA Base semilocal functional Dick & Fernandez-Serra (2021), Batch size 1, Iloss Number of steps to compute loss 3, Parameter EMA 0.995, Optimizer Adam β1 0.9 β2 0.999, Basis set 6-31G(d), Density fitting basis set weigend I, SCF iterations 15, Precycle XC functional LDA, Precycle iterations 15, Learning rate MD17 0.01 1+ 1 1000, 3BPA 0.01 1+ 1 1000, QM9 0.001 1+ 1 1000. |