Deep learning algorithms applied to computational chemistry.

Area of Science:

• Computational chemistry
• Molecular sciences
• Artificial intelligence in chemistry

Background:

• Deep learning (DL) is increasingly used in molecular sciences, demonstrating high performance and generalization.
• Existing DL models have limitations, and their comparative advantages/disadvantages are unclear for newcomers.
• A clear understanding of DL algorithms is crucial for advancing computational chemistry.

Purpose of the Study:

• To review deep learning algorithms applied to molecular challenges in computational chemistry.
• To provide a comprehensive categorization of conventional and geometric deep learning models.
• To analyze key features, open issues, and applications of these algorithms.

Main Methods:

• Categorization of DL algorithms into conventional and geometric approaches.
• Analysis of input descriptors, datasets, code availability, and task solutions.
• Review of research applications, trends, and future directions in molecular algorithm design.

Main Results:

• Detailed analysis of conventional and geometric DL algorithms, including their strengths and weaknesses.
• Information on datasets, libraries, computational costs (GPU, time), and optimization schemes.
• Identification of suitable algorithms for specific tasks and common data/input practices.

Conclusions:

• The review provides a structured overview of DL in computational chemistry, aiding algorithm selection.
• It serves as a reference for datasets, input data, and algorithm techniques.
• Insights into benefits and open issues support the development of novel computational chemistry systems.