Geometric Message Passing to utilize local and non-local information in 3D Graph Networks
Translated title
Geometric Message Passing to utilize local and non-local information in3D Graph Networks
Author
Svöðstein, Jákup Odssonur
Term
4. Term
Publication year
2021
Submitted on
2021-09-14
Pages
11
Abstract
Computer simulations of molecules work best when models capture both the relative 3D structure and long‑range interactions between atoms. In this work, we introduce a general framework, the deterministic point graph network (DPGN), that provides a unified way to work with 3D graphs across bonds, angles, torsions, and long‑range effects. We implement DPGN via a geometric message passing (GMP) scheme, where parts of a molecule exchange information based on their spatial relationships, and we assess each design choice through ablation studies. We validate the model by outperforming the predecessor MPNet on six properties in the QM9 dataset, achieving state‑of‑the‑art on one of them. We also demonstrate strong performance in molecular dynamics by reaching state‑of‑the‑art on 50% of the targets in the MD17 dataset.
Computersimuleringer af molekyler kan blive langt bedre, hvis modellerne forstår både den relative 3D‑struktur og langtrækkende interaktioner mellem atomer. I dette arbejde foreslår vi et generisk rammeværk, deterministic point graph network (DPGN), der giver en samlet måde at arbejde med 3D‑grafer på tværs af bindinger, vinkler, torsioner og langtrækkende effekter. Vi realiserer DPGN med et geometrisk message passing‑skema (GMP), hvor dele af molekylet udveksler information baseret på deres rumlige relationer, og vi viser værdien af de enkelte designvalg gennem ablationsstudier. Vi validerer modellen ved at overgå forgængeren MPNet på seks egenskaber i QM9‑datasættet, hvoraf én er state‑of‑the‑art. Derudover viser vi modellens styrke i molekylær dynamik ved at opnå state‑of‑the‑art på 50% af målene i MD17‑datasættet.
[This apstract has been rewritten with the help of AI based on the project's original abstract]