Machine learning presents a transformative approach for discovering novel compounds with desired properties, but more research is required to determine the optimal way to train machine learning models. Therefore, we have created the chemical autoencoder tool to allow users to joint train an autoencoder model on their data sets and generate PCA plots to better understand how different joint training scenarios affect the organization of the latent space. The latent space is created in the process of joint training these models and is a high dimensional space where the molecular structures are represented as continuous vectors (i.e. a list of numbers). With the tool, users can upload molecular structures in the form of SMILES data, corresponding property information, and a jointly trained autoencoder model. With these inputs, users can run joint training on multiple properties, even on data sets with missing property information. Users can also generate PCA plots from an existing model. The tool allows users to view 1D, 2D, and 3D PCA plots to explore the latent space organization of existing models. With these functionalities, users will be able to explore how to best train autoencoder models to organize the latent space.

• Python 3.6 and Anconoda 6
• Jupyter Notebook
• Hublib.UI 
• Tensorflow V.12
• Keras

• Network for Computational Nanotechnology

Nicolae C. Iovanac and Brett M. Savoie The Journal of Physical Chemistry A 2019 123 (19), 4295-4302 DOI: 10.1021/acs.jpca.9b01398

Researchers should cite this work as follows:

• Bryan Arciniega, Mackinzie S Farnell, Nicolae C Iovanac, Brett Matthew Savoie (2021), "Chemical Autoencoder for Latent Space Enrichment," https://nanohub.org/resources/chemae. (DOI: 10.21981/RGR3-Y309).

  BibTex | EndNote

1. data driven design
2. Jupyter notebooks
3. machine learning
4. material properties
5. materials science
6. principal component analysis