Rethinking end-to-end prediction of adsorption energies from a causal perspective
Keywords: Graph Neural Network, Material Discovery, Catalyst, AI for Science
Abstract: Adsorption energy is an important descriptor of catalytic activity in the field of catalysis, and significant efforts have been made to develop accurate predictive machine-learning models to replace expensive quantum chemistry calculations. Although it can be inferred by total energy predictions, research has mostly focused on the end-to-end prediction of adsorption energies due to the common belief that total energy is more challenging to predict than adsorption energy. In this study, we first analyzed the causal graph of adsorption energies and revealed that the indirect approach, which infers adsorption energy from total energy predictions, could provide better identifiability, leading to improved accuracy and generalization ability. We also improved the graph property normalization method for total energy prediction and achieved a halved Mean Absolute Error compared to direct adsorption energy prediction for the catalyst in-domain scenario. In the more challenging catalyst out-of-domain scenario, we found that the error primarily comes from predicting the individual energy of unseen catalyst atoms, and the error can be canceled when total energy predictions are used to infer adsorption energy. Consequently, our model achieves a MAE of approximately 0.2 eV for all tasks in the OC20 S2EF task, outperforming end-to-end models trained on datasets 50$\times$ larger. Given the evidence presented in this study, future research should prioritize the development of total energy models to enhance the accuracy and efficiency of machine-learning approaches in material discovery.
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5813
Loading