Skip to content

flawnson/Carbon_Molecule_Autoencoder-master

Repository files navigation

Project Carbon Coded

Description

A carbon molecule vanilla autoencoder. The latent layer is comprised of 7 nodes, and is a condensed representation of the learned features from the dataset of 12,000 data points. The latent layer is trained to understand the possible structures carbon is capable of forming. The data preprocessing is comprised of opening the data file, formatting the strings, and appending them into an array. The arrays are then turned into one hot encodings, each encoding representing another molecule. The one hot vectors are then fitted into the model.

Article

Here's an article I wrote about building this project, as well as how and why it failed https://towardsdatascience.com/building-a-carbon-molecule-autoencoder-21973e5f88b6

About

A Carbon molecule autoencoder neural network

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages