Variational Monte Carlo on a Budget — Fine-tuning pre-trained Neural Wavefunctions

Authors: Michael Scherbela, Leon Gerard, Philipp Grohs

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

Reproducibility Variable Result LLM Response
Research Type Experimental We extensively evaluate the accuracy, scalability and limitations of our base model on a wide variety of test systems.
Researcher Affiliation Academia Michael Scherbela University of Vienna michael.scherbela@univie.ac.at Leon Gerard* University of Vienna leon.gerard@univie.ac.at Philipp Grohs University of Vienna philipp.grohs@univie.ac.at
Pseudocode No No pseudocode or algorithm blocks were explicitly labeled or formatted as such in the paper.
Open Source Code Yes All code, configuration files, geometries, datasets and obtained energies are available on Git Hub under https://github.com/mdsunivie/deeperwin.
Open Datasets Yes A chemically diverse dataset with up to 100 molecules based on QM7-X [26]... To train our pre-trained wavefunctions we use two datasets, consisting of 18 and 98 disparate molecules... For the large scale experiment we used a stratified random sample of 250 molecules from the QM7 dataset [32].
Dataset Splits No The paper does not explicitly describe a separate validation dataset split with percentages or sample counts.
Hardware Specification Yes Overall we used 5k GPUhs (A100) for development and training of our base models, and another 5k GPUhs (A40) on evaluations and finetuning.
Software Dependencies No The paper mentions various software components such as 'py SCF [35]', 'e3nn library [36]', 'Adam optimizer [37]', 'RDKit [38]', 'ORCA [39]', 'Fermi Net codebase [40]', and 'KFAC [30] and their Python implementation [41]', but does not specify their exact version numbers for reproducibility.
Experiment Setup Yes A detailed description of the hyperparameter used in this work can be found below (cf. Tab. 2). Table 2: Hyperparameter settings used in this work.