learn_ml

Hands-on machine learning with Python: a sequence of classes followed by worked case studies, with a focus on materials-science and real (messy) data.

Setup

Requires Python ≥ 3.10.

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
jupyter notebook

Some case studies need extra, heavyweight dependencies that are commented out in requirements.txt (GPAW, Materials Project API, Google API client). Install them only if you intend to run those specific scripts.

Classes

Class	Contents
Classes/Class01_PythonReview	Python fundamentals for the rest of the course: variables, containers, control flow, functions, imports.
Classes/Class02_DataEngineering	NumPy and pandas from zero: arrays, vectorisation, DataFrames, missing data, one-hot encoding, outlier removal, text manipulation. Includes Lab_DataEngineering.ipynb exercises (with a solutions notebook).
Classes/Class03_ExploratoryDataAnalysis	EDA workflow on a deliberately dirty satellite database: dtype inspection, distributions, categorical counts, correlations, and what each plot tells you. eda.py is an advanced script version of a full EDA pipeline.
Classes/Class04_EnsembleLearning	Why ensembles work (bagging, boosting, voting); a CNN ensemble for image classification plus a tabular VotingClassifier/RandomForest example.

Case studies

Case study	Problem	Data	Methods
CaseStudy01_ClimateChange	Forecasting global temperature anomalies	Berkeley Earth (download or Data/GlobalTemperatures.csv)	SARIMAX, auto_arima, XGBoost, NN, naive baseline
CaseStudy02_BasicFeaturesForCrystals	Formation-energy prediction from elemental/structural descriptors	Materials Project (regeneration script provided)	XGBoost / RF / SVR regression
CaseStudy03_Molecules	HOMO–LUMO gap prediction from one-shot-DFT ("sDFT") spectral descriptors	Data/Molecules (GDB-9) + GPAW	XGBoost, feature-importance analysis
CaseStudy04_COVIDDeaths	Forecasting cumulative COVID-19 deaths	Data/time_series_covid19_deaths_global.csv	SARIMAX, auto_arima, XGBoost, NN
CaseStudy05_MalariaClassification	Parasitised vs healthy blood-cell images	Data/cell_images	CNN, hard/soft voting ensembles, patient-grouped evaluation
CaseStudy06_YoutubeComments	Downloading YouTube comments + intro NLP	YouTube Data API v3	TF-IDF + logistic regression
CaseStudy07_CrystalFeatures	Band-gap prediction from two descriptor families	committed CSVs (Materials Project + GPAW)	XGBoost, descriptor/label-leakage discussion

The script-based case studies (02, 03, 06, 07) each document their own pipeline — generate/prepare data → train model → evaluate — in their local README; the notebook case studies (01, 04, 05) are self-contained.

Data

Shared datasets live in Data/ and are documented in Data/README.md.