AI in Earth Sciences

Interactive lecture notes and practice exercises for machine learning in the geosciences

EarthSc 5757 — The Ohio State University — Joachim Moortgat

What this is

Open educational materials from EarthSc 5757: AI in Earth Sciences at The Ohio State University. The content covers machine learning foundations — from linear regression through convolutional neural networks — with examples and datasets drawn from the geosciences (seismic velocity prediction, isotope geochemistry, satellite land-cover classification, and more).

Everything is in Python (NumPy, scikit-learn, TensorFlow/Keras) and presented as interactive Jupyter notebooks that you can read here or run yourself.

Who this is for

✔ Students not enrolled in the course who want to learn ML with Earth Science applications
✔ Faculty developing their own courses who want to adopt or adapt individual modules
✔ Researchers looking for worked examples of ML applied to geoscience datasets
✔ Self-learners with some Python experience who want a structured path through ML fundamentals

Modules

Each module includes an overview (what you’ll learn and why), lecture notes (interactive Jupyter notebooks with theory, code, and visualizations), and practice exercises (hands-on problems with real data).

Introduction & Python Basics

Course overview, Jupyter notebooks, Python, NumPy, and Matplotlib fundamentals.

Univariate Linear Regression

Cost functions, gradient descent, and the optimization framework that underlies all subsequent methods.

Multivariate Regression

Multiple features, polynomial models, feature scaling — applied to seismic velocity data and housing prices.

Classification

Logistic regression for binary and multiclass problems — including handwriting digit recognition.

Data Preprocessing & Pipelines

Feature scaling, encoding, imputation, and scikit-learn Pipelines for reproducible workflows.

Cross-Validation & Regularization

Bias vs. variance, k-fold CV, Ridge, Lasso, Elastic Net, and GridSearchCV.

Artificial Neural Networks

Perceptrons, hidden layers, backpropagation — an ANN coded from scratch.

Convolutional Neural Networks

Convolutions, pooling, CNN architectures — applied to satellite land-cover classification and image segmentation.

Acknowledgments

The early development of this course was inspired by Andrew Ng’s Machine Learning course (Stanford / Coursera). The original Matlab/Octave materials were translated to Python and have been substantially reworked over several years to focus on Earth Science applications, geoscience datasets, and modern Python tooling (scikit-learn, TensorFlow/Keras).

BuckAI Observatory · The Ohio State University · School of Earth Sciences