Skip to content

1000 — Machine learning

This chapter introduces machine learning as an engineering tool for propulsion problems: not as “AI magic”, but as function approximation under constraints.

In propulsion and aircraft system studies, we often need fast surrogates for:

  • cycle calculations,
  • component maps,
  • mission-level performance evaluation,
  • design space exploration.

A feed-forward neural network (FFNN) is one of the simplest ways to learn a mapping:

\[ \mathbf{x} \;\mapsto\; \mathbf{y} \]

where:

  • \(\mathbf{x}\) are operating/design inputs,
  • \(\mathbf{y}\) are outputs predicted by a high-fidelity or expensive model.

The goal is to understand:

  • what an FFNN can and cannot learn,
  • what “good performance” means beyond a low loss value,
  • how to evaluate generalization and failure modes.

Exercises in this chapter

1000.1 — FFNN regression on engine data (PyTorch)

Train a simple neural network to predict engine outputs from input parameters using a dataset stored in .dat files.

➡️ Exercise page →


Learning outcomes

After completing this chapter, you should be able to:

  • Train and evaluate a basic FFNN regression model
  • Explain the difference between training error and generalization error
  • Diagnose underfitting vs overfitting
  • Recognize the role of scaling/normalization in regression
  • Produce meaningful plots (parity plots, residuals) to validate a surrogate
  • Discuss how this kind of model could be used (and misused) in system studies