Multi-Channel 1D PolyConvNet

This project is part of ALCF AI Hackathon on molecular dynamics hosted by Argonne Leadership Computing Facility (ALCF).

We proposed a multi-channel 1D convolutional neural network for predicting the lamellar period of copolymers based on the sequence of beads.

Team Members: Ruijie Zhu (Northwestern), Kastan Day (UIUC), Aria Coraor (UChicago), Seonghwan Kim (UIUC), Jiahui Yang (Northwestern)

Github link

The neural network takes in 4 types of features as input:

1. Sliding Window Features

used to capture the activation of polymer sequence (29-dimensional)

2. Kernels

• Exponential kernel: used to capture the interactions at two ends (30-dimensional)

• Cosine kernel: used to capture the periodicity of sequence (15-dimensional)

3. VAE Features (concatenated after Conv1d Layer)

generated using the Variational Autoencoder model (4-dimensional)

4. Interaction Parameters (concatenated after Conv1d Layer)

Our team won the 1st place of the hackathon. Shout out to all my teammates!