Time frequency representations and deep convolutional neural networks: a recipe for molecular properties prediction

From National Research Council Canada

DOI	Resolve DOI: https://doi.org/10.1109/CCECE53047.2021.9569072
Author	Search for: Tchagang, Alain B.¹; Search for: Valdes, Julio J.¹
Affiliation	National Research Council Canada. Digital Technologies
Format	Text, Article
Conference	2021 IEEE Canadian Conference on Electrical and Computer Engineering (CCECE), Sept 12-17, 2021, Held Virtually
Subject	convolutional neural networks; Coulomb matrix; density functional theory; molecular forward design; time-frequency representations
Abstract	In recent years, Quantum Mechanics (QM) has been combined with Machine Learning (ML) algorithms to speed up the design of molecules, drugs and materials. These paradigms known as QM↔ML have been successful in providing the precision of QM at the speed of ML. In this work, we show that by integrating well-known signal processing (SP) techniques in the QM↔ML pipeline, we obtain a powerful methodology (QM↔SP↔ML) that can be used for representation, visualization and molecular properties predictions. Tested on the benchmark QM9 dataset, the new QM↔SP↔ML framework is able to predict the properties of molecules with a mean absolute error below acceptable chemical accuracy, and yield better or similar results compared to other ML state-of-the-art techniques described in the literature.
Publication date	2021-10-26
Publisher	IEEE
In	2021 IEEE Canadian Conference on Electrical and Computer Engineering (CCECE) (26 October 2021): 1–5.
Language	English
Peer reviewed	Yes
Export citation	Export as RIS
Report a correction	Report a correction (opens in a new tab)
Record identifier	1237329f-fe08-4c94-bca6-65079da70938
Record created	2021-10-28
Record modified	2021-10-29

Date modified:: 2025-12-13