Material Prediction and Design

Finding the relationship between the composition/structure and properties of materials is a holy grail of material science. Various models ranging from QM methods on the atomic scale to FEM on the macroscopic scale have been developed in the past century. More recently, ML-based approaches have gained momentum.

Here, we provide a toy dataset of alloys toy_alloy_data.csv. The first 12 columns correspond to the proportion of 12 elements in the alloy, whose sum is one. The last 6 columns correspond to 6 properties.

It is generally not difficult to build a model to predict properties from composition. We show how to train a simple neural network model for the composition-to-property problem using PyTorch in DeepVerse_Challenge_1.ipynb.

Challenge 1
- 1. Can you try to improve it?
- 2. In case of missing data, which is typical of experiment studies, how would you deal with that? toy_alloy_data_with_missings.csv
- 3. More importantly, can you find a way to predict compositions from a given set of desired properties? You are not restricted to NN models only. However, note the predicted compositions should all be positive and sum to one.


Predict elasticity of crystals

The Materials Project (https://materialsproject.org/) is an open-access database offering material properties, where the elastic properties of 13822 compounds have been reported.

Challenge 2
- 1. Develop and train a machine learning model which is able to predict stiffness tensor from the composition of crystals only.
- 2. Develop and train another machine learning model which is able to predict stiffness tensors from the composition and the structure of crystals. (One may want to refer to the so-called crystal graph neural networks https://tony-y.github.io/cgnn/)
- 3. Compare the accuracy of the above two models and explain why one is better than the other.
- 4. Repeat the same procedure to predict piezoelectric tensors of crystals.