AdsorbDiff: Adsorbate Placement via Conditional Denoising Diffusion

Authors: Adeesh Kolluru, John R. Kitchin

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our findings demonstrate an acceleration of up to 5x or 3.5x improvement in accuracy compared to the previous best approach. Given the novelty of this framework and application, we provide insights into the impact of pre-training, model architectures, and conduct extensive experiments to underscore the significance of this approach.
Researcher Affiliation Academia 1Department of Chemical Engineering, Carnegie Mellon University. Correspondence to: Adeesh Kolluru <akolluru@andrew.cmu.edu>, John R. Kitchin <jkitchin@andrew.cmu.edu>.
Pseudocode Yes For the initial coordinates of adsorbate, we select a random point on the slab. This point is considered as the center-of-mass of the adsorbate in fractional coordinates. We then convert from fractional coordinates to real coordinates and perform a reverse diffusion process to get to the lowest energy site (as shown in Algorithm 1).
Open Source Code Yes The code is open-sourced with MIT License1. 1https://github.com/Adeesh Kolluru/ Adsorb Diff
Open Datasets Yes We utilize two publicly available datasets for this work OC20-Dense (Lan et al., 2023) and OC20 (Chanussot et al., 2021).
Dataset Splits Yes We split the ID data into 80/20 ratios for training the diffusion model and validating the sampling process.
Hardware Specification Yes We perform all of our training on 2 48GB A6000 GPUs.
Software Dependencies No The paper mentions using specific algorithms like L-BFGS and models like Equiformer V2, Gem Net-OC, and Pai NN, but it does not provide specific version numbers for software libraries or dependencies like Python, PyTorch, etc.
Experiment Setup Yes The hyperparameters of L-BFGS are mentioned in Table 4. We set a max step size for sampling and stop early if the movement of atoms converges. These hyperparameters are shown in Table 5. Maximum learning rate was kept to be 4.e-4 and 1.e-4 for pretraining and finetuning runs respectively utilizing Adam W optimizer. For each model architecture, maximum batch size was utilized that could fit on the GPU.