Towards Combinatorial Generalization for Catalysts: A Kohn-Sham Charge-Density Approach
Authors: Phillip Pope, David Jacobs
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | On a new dataset of bulk catalysts with charge densities, we show density models can generalize to new structures with combinations of elements not seen at train time, a form of combinatorial generalization. We show that over 80% of binary and ternary test cases achieve faster convergence than standard baselines in Density Functional Theory, amounting to an average reduction of 13% in the number of iterations required to reach convergence, which may be of independent interest. Our results suggest that density learning is a viable alternative, trading greater inference costs for a step towards combinatorial generalization, a key property for applications. |
| Researcher Affiliation | Academia | Phillip Pope University of Maryland, College Park pepope@cs.umd.edu David Jacobs University of Maryland, College Park dwj@umd.edu |
| Pseudocode | No | The paper does not include pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | No | The paper does not provide a direct link or explicit statement that the source code for the methodology developed in *this paper* is being released. It mentions: "Our use of open-source DFT solvers is one notable step towards this goal over previous approaches [8]", which refers to the tools they used, not their own implementation code. |
| Open Datasets | No | Although the paper describes creating a "new large-scale DFT dataset" and details its characteristics, it does not provide concrete access information (e.g., a link, DOI, or repository name) for this newly generated dataset to be publicly accessed. It references existing tools and projects (Materials Project, Open Catalyst Project) but not the specific dataset generated for this work. |
| Dataset Splits | Yes | We then partition the train/val/test splits as follows. First we assign all unary catalysts remaining to the train set. Then we randomly sample binaries with oxidation states not already represented in the unaries. Next, we remove oxidation states not present in the training set from the remaining binary and ternary candidates. This preprocessing step performed to ensure all states in the val/test splits were represented in the training set, e.g. training on H+ cannot be expected to generalize to H . Finally we assign a few binaries to the validation split and the remaining binaries and ternaries to the testing split. Finally we validate that train and test splits are disjoint by element combinations, i.e. no combination of elements in training set is repeated in the test splits. We report the number of structures and density samples in Table 1. (Table 1: Nstructures Nsamples Train (unary) 47 3.8M Train (binary) 392 69M Val (binary) 4 300k Test (binary) 360 72M Test (ternary) 1116 380M) |
| Hardware Specification | Yes | Regarding compute, for training we used approximately 2000 GPU hours for both replicates across all GPUs. We use NVIDIA A5000 cards with 24 GB of memory. |
| Software Dependencies | No | The paper mentions using "Quantum Espresso" and "Aii DA" for dataset generation and simulations, and "pymatgen" for preprocessing. It also mentions "ASE" for band gap computation. However, specific version numbers for these software packages or any other libraries are not provided, which is necessary for reproducibility. |
| Experiment Setup | Yes | We train a reduced sized Spherical Channel Network (SCN) for density prediction. We reduce the number of interactions to 10 and hidden channels to 256, resulting in 35M parameters, which is smaller than state of the art models [55]. We use interaction cutoff to 6.0Å, and a maximum number of neighbors per node of 12. We list full hyperparameters in the Supplementary. We train two replicates of SCNs for 420k steps with batch size 45 across 8 GPUs. |