Machine Learning for Materials (MATE70026)

To follow the tutorials in this lecture, you need to install scikit-image for image processing and torchvision for computer vision datasets.

!pip install scikit-image --quiet
!pip install torchvision --quiet

You can explore available datasets in matminer using matminer.datasets.get_available_datasets(). To get detailed information about a specific dataset, use matminer.datasets.dataset_retrieval.get_all_dataset_info(). Once identified, load the dataset into a pandas DataFrame using load_dataset() from matminer.datasets.dataset_retrieval.

import matminer
from matminer.datasets.dataset_retrieval import load_dataset

# List available datasets
matminer.datasets.get_available_datasets(print_format='low')

# Get detailed info for a specific dataset
print(matminer.datasets.dataset_retrieval.get_all_dataset_info('matbench_log_kvrh'))

# Load the dataset
df = load_dataset('matbench_log_kvrh')

To view the course website locally, you must use jupyter-book. First, install the jupyter-book package via pip, then navigate to the repository root and execute the build command.

pip install -U jupyter-book

# Navigate to the repository directory
cd mlformaterials

# Build the local site
jupyter-book build .

The course materials are available as an online resource or can be downloaded for local use. You can browse individual files in the repository or download the entire course folder as a ZIP file to work offline.

https://github.com/aronwalsh/MLforMaterials/archive/master.zip

For developers transitioning from physics or chemistry to computational workflows, the following resources are recommended:

• Python for Physicists: lucydot.github.io/python_novice
• Python for Chemists: pythoninchemistry.org
• AI Assistance: Phind AI helper

To follow the workflows in this lecture, install the following libraries using pip. These include matminer for materials featurization, smact for charge-balanced composition screening, plotly for visualization, and multiprocess for parallel computing.

!pip install matminer --quiet
!pip install smact==3.2.0 --quiet
!pip install plotly --quiet
!pip install multiprocess --quiet

To ensure a model generalizes to unseen data and is not overfitted, use $k$-fold cross-validation. Using sklearn.model_selection.KFold, you can split the data into $k$ subsets, training on $k-1$ and validating on the remaining one. Use cross_val_score to evaluate specific metrics like neg_mean_absolute_error or r2.

from sklearn.model_selection import KFold, cross_val_score

# Setup KFold
cv = KFold(n_splits=5, shuffle=True, random_state=42)

# Evaluate using MAE and R2
mae_scores = cross_val_score(rf, X, y, cv=cv, scoring='neg_mean_absolute_error')
r2_scores = cross_val_score(rf, X, y, cv=cv, scoring='r2')

print(f'mean MAE = {np.mean(np.abs(mae_scores)):.3f}')
print(f'mean r^2 = {np.mean(r2_scores):.3f}')

To follow the tutorials in this repository, you need to install matminer for materials informatics and seaborn for statistical visualization. You can check your numpy version using pip show numpy to ensure compatibility.

!pip install matminer --quiet
!pip install seaborn --quiet
!pip show numpy

Bayesian Optimisation (BO) is a technique for optimising complex, expensive-to-evaluate functions (like physical experiments).

Core Components:

1. Surrogate Model: Typically a Gaussian Process (GP) that approximates the true objective function $O(x)$. It captures both the mean $\mu(x)$ and the uncertainty $\sigma(x)$ at any given point.
2. Acquisition Function $\alpha(x)$: A mathematical function that decides where to sample next by balancing exploration (sampling in uncertain regions) and exploitation (sampling in regions known to be promising). Common functions include:
   • Probability of Improvement (PI)
   • Expected Improvement (EI)
   • Upper Confidence Bound (UCB)

In materials research, BO is used to find optimal parameters like chemical composition or processing conditions without exhaustive trial-and-error.

Principal Component Analysis (PCA) can be used to reduce high-dimensional featurized data into 2D or 3D space for visualization.

Workflow:

1. Standardize features using sklearn.preprocessing.StandardScaler.
2. Initialize sklearn.decomposition.PCA with n_components (e.g., 2 or 3).
3. Use .fit_transform(X) to obtain the reduced coordinates.
4. Plot using matplotlib or plotly.

from sklearn.decomposition import PCA
from sklearn.preprocessing import StandardScaler

# 1. Standardize
X_standardised = StandardScaler().fit_transform(X)

# 2. & 3. Reduce to 2D
pca = PCA(n_components=2)
reduced_data = pca.fit_transform(X_standardised)

# 4. Check variance
print(f"Cumulative variance: {pca.explained_variance_ratio_.sum()}")