/pred-membrane

Primary LanguagePython

MembraneAI: Machine Learning Guided Polyamide Membrane with Exceptional Solute-Solute Selectivity and Permeance

This repo is a support for our publication Machine Learning Guided Polyamide Membrane with Exceptional Solute-Solute Selectivity and Permeance by Hao Deng, Zhiyao Luo, Joe Imbrogono, Tim M Swenson, Zhongyi Jiang, Xiaonan Wang, and Sui Zhang. A special thanks goes to Yaotian Yang for his generous help on sorting out this repo.

General info

Herein, we present a data-driven strategy to guide the design of polyamide membrane by developing a novel artificial neural network (ANN) based multi-task machine learning (ML) model with skip connections and selectivity regularization. We used limited sets of laboratory collected data to obtain satisfactory model performance over four iterations of an ever-expanding dataset incorporating with human expert experience in the online learning process. We then fabricated four membranes under the guidance of model, and they all exceed the present upper bound for mono/divalent ion selectivity and permeance of the polymeric membranes. Moreover, we obtained mechanistic insights into the membrane design through feature analysis of the model. Our work demonstrates that ML approach represents a paradigm shift for high-performance polymeric membranes design.

Table of contents

Setup

To run the code, first ensure that the Python Interpreter version is python3.8 or another appropriate version, and then install the appropriate module using the pip install -r requirements.txt .

use

To simply run by the train-test manner, run train_torch.py directly

Alternatively, run train_validate.py to get the optimal hyperparameters on the train-val dataset, and retrain models by running train_torch.py with the optimized hyperparameters. image

file structure

│ 
├─utils.py                  # helper functions for data split and training/testing
│
├─cross_validation          # Use k-fold test method of cross validation to evaluate the model performance and optimize model overparameters
│      validate_ensemble.py # Validate emsemble models such as random forest and SVR
│      validate_nn.py       # Validate ANN-based models such as native MLP and our model
│      __init__.py
│      
├─data
│       # datasets for 5 iterations, named by the number of unique samples
│      
├─train_model
│      torch_models.py       # Different MLP models are defined and written
│      train_torch.py        # The body function for training, which can be run directly
│      __init__.py
│      
└─visu                       # Visualization of data results and data characteristics
        draw_pdp.py
        eval_multiregression.py
        feature_importance.py
        plotting.py
        __init__.py