SKM 2023 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

HL: Fachverband Halbleiterphysik

HL 6: Focus Session: Frontiers of Electronic-Structure Theory III (joint session O/HL)

HL 6.8: Vortrag

Montag, 27. März 2023, 12:45–13:00, TRE Ma

Analysis of Batching Methods in Graph Neural Network Models for Materials Science — •Daniel Speckhard, Tim Bechtel, Jonathan Godwin, and Claudia Draxl — Humboldt-Universität zu Berlin, Physics Department and IRIS Adlershof, Berlin, Germany

Graph neural network (GNN) based models have shown promising results for materials science [1]. These models often contain millions of parameters, and like other big-data based models, require only a portion of the entire training dataset to be fed as a mini-batch to update model parameters. The effect of batching on the computational requirements of training and model performance has been thoroughly explored for neural networks [2] but not yet for GNNs. We explore two different types of mini-batching methods for graph based models, static batching and dynamic batching. We use the Jraph library built on JAX to perform our experiments where we compare the two batching processes for two data-sets, the QM9 dataset of small molecules and the AFLOW materials database [3]. We show that dynamic batching offers significant improvements in terms of computational requirements for training. We also present results on the effect of the batch size and batching method on model performance.

[1] T. Xie et al., Physical Review Letters, 120, 14 (2018).

[2] M. Li et al., Proceedings of the 20th ACM SIGKDD (2014).

[3] S. Curtarolo et al., Comp. Mat. Science, 58, 227-235 (2012).