We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations

DBLP - CS Bibliography


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: PhAST: Physics-Aware, Scalable, and Task-specific GNNs for Accelerated Catalyst Design

Abstract: Mitigating the climate crisis requires a rapid transition towards lower carbon energy. Catalyst materials play a crucial role in the electrochemical reactions involved in a great number of industrial processes key to this transition, such as renewable energy storage and electrofuel synthesis. To reduce the amount of energy spent on such processes, we must quickly discover more efficient catalysts to drive the electrochemical reactions. Machine learning (ML) holds the potential to efficiently model the properties of materials from large amounts of data, and thus to accelerate electrocatalyst design. The Open Catalyst Project OC20 data set was constructed to that end. However, most existing ML models trained on OC20 are still neither scalable nor accurate enough for practical applications. Here, we propose several task-specific innovations, applicable to most architectures, which increase both computational efficiency and accuracy. In particular, we propose improvements in (1) the graph creation step, (2) atom representations and (3) the energy prediction head. We describe these contributions and evaluate them on several architectures, showing up to 5$\times$ reduction in inference time without sacrificing accuracy.
Comments: Accepted at the NeurIPS 2022 AI for Accelerated Materials Design Workshop
Subjects: Machine Learning (cs.LG); Computational Physics (physics.comp-ph)
Cite as: arXiv:2211.12020 [cs.LG]
  (or arXiv:2211.12020v1 [cs.LG] for this version)

Submission history

From: Victor Schmidt [view email]
[v1] Tue, 22 Nov 2022 05:24:30 GMT (155kb,D)

Link back to: arXiv, form interface, contact.