Generalizing Neural Wave Functions

Authors: Nicholas Gao, Stephan Günnemann

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In our experiments, we find Moon converging in 4.5 times fewer steps to similar accuracy as previous methods or to lower energies given the same time. Further, our analysis shows that Moon s energy estimate scales additively with increased system sizes, unlike previous work where we observe divergence. 5. Experiments Here, we analyze Globe and Moon across a variety of different experimental settings.
Researcher Affiliation Academia Nicholas Gao 1 Stephan G unnemann 1 1Department of Computer Science & Munich Data Science Institute, Technical University of Munich, Germany. Correspondence to: Nicholas gao <n.gao@tum.de>.
Pseudocode Yes Algorithm 1 Orbital localization
Open Source Code Yes 1Source code: https://www.cs.cit.tum.de/daml/globe/
Open Datasets Yes For the six-element hydrogen chain, we use the pretraining geometries from Scherbela et al. (2022), and for the ten-element hydrogen chain, we use the geometries from Motta et al. (2017). For testing dissimilar structures, we use the same distances for nitrogen as in Pfau et al. (2020). For ethene, we use the evaluation structures from Scherbela et al. (2022). In our transferability experiment, we take the six-element and ten-element hydrogen chain as well as the methane, and ethene structures and energies from Scherbela et al. (2022). The cyclobutadiene structures are from Lyakh et al. (2012). For benzene, we reuse the same geometry from Ren et al. (2022) as previous works.
Dataset Splits No The paper describes training but does not explicitly provide details on training/validation/test splits, specific validation sets, or cross-validation methodologies.
Hardware Specification Yes All experiments ran on 1 to 4 Nvidia A100 GPUs depending on the system size.
Software Dependencies No The paper mentions implementing experiments in JAX and using Py SCF, but it does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes Table 2. Default hyperparameters. Hyperparameter Value Pretraining Steps 1e4 Basis STO-6G Method RHF Optimization Steps 6e4 Learning rate 0.1 / 1+ t 100 Batch size 4096 Damping 1e-4 σ[EL] Local energy clipping 5 Max grad norm 1 CG max steps 100 Moon Hidden dim 256 E-E int dim 32 Layers 4 Activation Si LU Determinants 16 Jastrow layers 3 Filter hidden dims [16, 8] Reparametrization Embedding dim 128 MLP layers 4 Message dim 64 Layers 3 Activation Si LU Filter hidden dims [64, 16]