NO. 41 · Computational & Data

Machine Learning for Geoscientists

Models that respect the geology and the test set. From problem framing and gradient descent through the classical toolkit to CNNs, with the honesty checklist that keeps a model out of trouble.

You can frame a geoscience problem for learning, train and evaluate the classical models and a first neural network in Python, diagnose overfitting when it happens, and report results a skeptical colleague would accept.

18 competencies · 6 interactive widget challenges · 10 to 16 hours of guided study

For geoscientists adding machine learning to their toolkit

Draws on: Applied Machine Learning for Geoscientists Statistics and Data Science for Researchers

Framing

Framing the problem

Forward versus inverse, prediction versus explanation: two distinctions that decide whether your model answers the question anyone asked.

Reads: Why ML for Geoscience? Problem Framing and Inference The ML mindset: prediction vs explanation

The geoscience data pipeline

From sensors to petabytes, the pipeline is where most project time actually goes; respect it before the modeling starts.

Reads: From Sensors to Petabytes: The Geoscience Data Pipeline

The optimizer and the statistics

Gradient descentwidget challenge

One loop trains everything from logistic regression to deep networks; walk it by hand once and nothing downstream is mysterious.

Reads: Gradient Descent and Sampling Strategies

The statistics underneath

Distributions, moments, and correlation are the load-bearing statistics of every model you are about to train; the lab makes them muscle memory.

Reads: Background Statistics Statistics Practice with Python

The classical toolkit

Linear and logistic regressionwidget challenge

The baselines that every fancier model must beat, and the confusion matrix and ROC curve that referee the contest.

Reads: Linear and Logistic Regression Regression Practice with Python

Nearest neighbors

KNN is the honest lazy model: no training, all memory, and a live demonstration of why feature scaling matters.

Reads: Nearest Neighbor and KNN KNN with Geoscience Data

Decision trees

Trees split the world the way interpreters already think, which is exactly why they overfit with such confidence.

Reads: Decision Trees Decision Tree Practice with Python

Random forestswidget challenge

Averaging many overfit trees into one calibrated committee is the single best default model in applied geoscience.

Reads: Random Forest Random Forest Practice

Naive Bayes

The independence assumption is obviously false and the model works anyway; understanding why teaches you more than most courses.

Reads: Naive Bayes Naive Bayes Practice

Features and failure modes

Feature engineering

Scaling, encoding, and domain features move models more than algorithm choice does; this is where geoscience knowledge pays.

Reads: Feature Engineering Feature Engineering Practice

Overfitting and the bias-variance tradewidget challenge

Every ML failure you will ever debug is somewhere on this curve; learn to point at where.

Reads: Overfitting, Underfitting, Variance, and Bias

Dimensionality reduction

PCA is how you look at a hundred-dimensional dataset without lying to yourself about what is really varying.

Reads: Dimensionality Reduction Dimensionality Reduction Practice

Deep learning

Neurons and a first network

A perceptron is logistic regression with ambition; stacking them is where representation learning starts.

Reads: Perceptrons and Neurons: A Simple NN Model Building a Simple Neural Network

Convolutional networkswidget challenge

Convolution is the same operation your seismic data was born from, which is why CNNs read geophysical images so naturally.

Reads: Deep Learning: Convolutional Neural Networks CNN Practice

Autoencoders

Compress, then reconstruct: the autoencoder is unsupervised learning earning its keep on unlabeled geoscience data.

Reads: Auto-encoders Auto-encoder Practice

Recurrent networks

Logs, time series, and anything with memory: RNNs carry state the way the subsurface carries history.

Reads: Recurrent Neural Networks

Ship it honestly

Reporting a result people can trustwidget challenge

A model is a claim; the reporting checklist is what makes the claim auditable instead of promotional.

Reads: Reporting an ML result so the reader can trust it

When not to use ML

The most senior skill in the field is declining to model; knowing the cases where a physical law or a lookup table wins.

Reads: When NOT to use ML