Engineering AI method selector
Choose a method only after stating the engineering question, evidence available, and decision consequence. This page is a routing aid, not a model-recommendation engine.
| If the engineering task is… | Start with… | Then verify… | Current route |
|---|---|---|---|
| Estimate a continuous quantity | Regression and a physically meaningful baseline | Units, residuals, operating-range coverage, and held-out error | Module I |
| Sort images, signals, or records into categories | Classification and a transparent threshold/rule baseline | Group/site/device split, class imbalance, false-alarm cost, and confusing cases | Module II |
| Find structure without a supplied label | PCA/SVD or clustering | Train-only representation, reconstruction, stability, and physical interpretation | Module III |
| Locate or measure regions/objects in imagery | Segmentation or object detection | Spatial labels, condition-held-out performance, visual errors, and measurement uncertainty | Module IV |
| Choose actions over time | A baseline controller before reinforcement learning | State/action units, safety constraints, reward design, perturbations, and sim-to-real gap | Module V |
| Create synthetic samples or propose candidates | Constraint screen and non-generative comparison | Provenance, novelty/memorization, task utility, physical feasibility, and fabrication limits | Module VI |
| Forecast a dynamic signal or detect degradation | Persistence/AR baseline before a sequence model | Chronological split, horizon-dependent error, drift, missingness, and decision cost | Module VII |
| Combine governing equations with data | Analytical/numerical solution before a PINN | Boundary/initial conditions, residuals, inverse uncertainty, and model-form error | Module VIII |
| Delegate a bounded workflow to an AI agent | Written workflow/specification before tool access | Permissions, provenance, tool-output checks, acceptance tests, and human responsibility | Module IX |
If the problem does not have a defensible target, baseline, data/right-to-use statement, or validation route, do not select a more complex model. Reframe the engineering question first.
For discipline-specific examples, use the application-studio roadmap.