1. Supervised Learning — learning from labelled data
Supervised learning is by far the most widely used type in materials science. The word supervised means that a human (or a DFT calculation) has already provided the correct answers — called labels — for each training example.
Given a training set {(x₁, y₁), (x₂, y₂), …, (xN, yN)}, learn a function f such that f(x) ≈ y for new, unseen inputs x.
Here x is the feature vector (structure) and y is the label (property).
Two sub-types of supervised learning
| Sub-type | Output type | Materials example | Algorithms |
|---|---|---|---|
| Regression | Continuous number | Predict band gap (eV), formation energy (eV/atom), bulk modulus (GPa) | Linear regression, SVR, neural networks, ALIGNN |
| Classification | Discrete category | Metal or insulator? Stable or unstable? Magnetic or non-magnetic? | Logistic regression, SVM, random forest, CGCNN |
In your daily work, DFT acts as the "supervisor" — it provides the ground-truth labels. You run hundreds of WIEN2k or VASP calculations, collect the results (Eg, Ef, magnetic moment…), and these become the labelled training data for your ML model. DFT is expensive but reliable; ML is fast but needs DFT's guidance.
The workflow you already built in your MxCy pipeline — predicting electronic and magnetic properties from composition — is a perfect example of supervised regression: DFT values as labels, physics-motivated features as inputs.
2. Unsupervised Learning — finding hidden structure
Unsupervised learning works without any labels. The algorithm receives only the input data x and must discover meaningful patterns on its own. There is no "correct answer" to compare against.
Given a dataset {x₁, x₂, …, xN} with no labels, find structure, patterns, or a compact representation of the data.
Main unsupervised tasks in materials science
| Task | What it does | Materials example |
|---|---|---|
| Clustering | Groups similar materials together automatically | Discover that your 500 compounds naturally fall into 4 structural families |
| Dimensionality reduction | Compresses many features into 2–3 for visualisation | PCA/t-SNE map of 50-feature crystal descriptors → 2D plot showing material families |
| Anomaly detection | Identifies unusual data points | Flag DFT calculations with suspiciously high energies — likely convergence errors |
| Generative models | Learns to generate new examples | Generate new stable crystal structures not in the training database |
Without labels, it is harder to measure how "good" an unsupervised model is. You cannot compute MAE or R². Instead you use domain knowledge: do the discovered clusters make physical sense? This requires your expertise as a materials scientist.
3. Reinforcement Learning — learning by trial and error
Reinforcement learning (RL) is the most different from the other two. There is no dataset at all — instead, an agent interacts with an environment, takes actions, and receives rewards or penalties.
An agent in state s takes action a, transitions to state s', and receives reward r. The goal: learn a policy π(s) → a that maximises cumulative reward over time.
| RL concept | Materials science analogy |
|---|---|
| Agent | The optimisation algorithm |
| Environment | The materials synthesis or simulation space |
| State | Current synthesis conditions (temperature, pressure, composition) |
| Action | Change a synthesis parameter |
| Reward | +1 if material is stable and has target property, −1 otherwise |
| Policy | The strategy: given current conditions, what to change next |
RL is the least mature of the three for materials applications. It is computationally expensive (requires many environment interactions) and difficult to set up. However, it is promising for inverse design — finding synthesis conditions that produce a material with target properties. For beginners, focus on supervised learning first.
How to choose the right type for your problem
Decision guide
e.g. band gap
e.g. metal?
find groups
create new
Complete algorithm overview
| Algorithm | Type | Task | Materials use case |
|---|---|---|---|
| Linear Regression | Supervised | Regression | Predict formation energy from composition |
| Logistic Regression | Supervised | Classification | Metal vs insulator prediction |
| Random Forest | Supervised | Both | Band gap, stability, property screening |
| Support Vector Machine | Supervised | Both | Phase classification, property prediction |
| Neural Network | Supervised | Both | General property prediction |
| CGCNN / ALIGNN | Supervised | Regression | DFT property prediction from crystal graph |
| K-Means | Unsupervised | Clustering | Group crystals by structural similarity |
| PCA | Unsupervised | Dim. reduction | Visualise high-dimensional material space |
| t-SNE / UMAP | Unsupervised | Dim. reduction | 2D maps of materials databases |
| VAE / GAN | Unsupervised | Generative | Generate new hypothetical crystal structures |
| Bayesian Optimisation | RL-adjacent | Optimisation | Active learning for synthesis optimisation |
Start with supervised learning — specifically regression. You already have DFT data (labels) and crystal structures (features). This maps directly onto your existing computational workflow. Posts 5–8 cover the core supervised algorithms in detail, always with materials science examples.
My name is Dr Abderrahmane Reggad. I'm a 54 year old and I am a researcher in materials' sciences.