StandardInternational professional-learning synthesis · ABET 2026–27 · IEA GAPC v4 · ASCE ethics and AI responsibility · NIST AI RMF 1.0
Source3 source files
Review stateTechnical and publication gates pending
After this chapter, you should be able to
- Distinguish physics features, constraints and residual learning.
- Build and test a bounded surrogate.
- Compare interpolation, edge and out-of-domain error.
- Identify model-form and boundary-condition risk.
- Reject unsupported optimisation and digital-twin claims.
Engineering context and methodSource §Lesson 12 · Engineering context and method · RES-09 · RES-01 · CUR-02
Physics may enter inputs, architecture, loss, post-processing or residual correction. This can improve plausibility or data efficiency but inherits wrong equations, boundary conditions and calibration. A surrogate approximates an authoritative calculation only within a declared and validated domain. An operational digital twin needs asset linkage, state updating, uncertainty, validation and a defined decision purpose—not merely a static model.
Verified worked exampleSource §Lesson 12 · Verified worked example · RES-09 · RES-01 · CUR-02
Cantilever deflection and surrogate error
q = 5,000 N/m, L = 3 m, E = 30×109 Pa and I = 0.004 m4; surrogate predicts 0.440 mm.
- Equation
δ = qL4/(8EI)
uniform-load cantilever assumption - Exact
5000×34 / (8×30×109×0.004)
0.000421875 m = 0.421875 mm - Error
|0.440−0.421875| / 0.421875 × 100
4.296%
Result. Absolute error = 0.018125 mm and relative error ≈ 4.30% at one point. This does not validate the domain or the structural assumptions.
Practical lab · 5 h lesson effortSource §Lesson 12 · Practical lab · 5 h lesson effort · RES-09 · RES-01 · CUR-02
- Generate reviewed synthetic cases within declared q, L, E and I ranges.
- Train a small surrogate and compare against the exact equation on an independent grid.
- Map maximum and 95th-percentile error, especially near domain edges.
- Test beyond range and implement a visible abstention rule.
Failure modes to investigateSource §Lesson 12 · Failure modes to investigate · RES-09 · RES-01 · CUR-02
- Wrong units or boundary conditions.
- Sparse coverage at domain edges.
- Physics penalty assumed exact.
- Optimiser exploits surrogate error.
- Static, uncalibrated model called a digital twin.
Knowledge checksSource §Lesson 12 · Knowledge checks · RES-09 · RES-01 · CUR-02
| Question | Answer rationale |
|---|---|
| Does physics-informed mean verified? | No; the physics representation and implementation can be wrong. |
| What is a surrogate domain? | The input and condition region supported by validation evidence. |
| Why test edges? | Approximation and extrapolation errors often grow there. |
| Can an optimiser trust the surrogate blindly? | No; it may exploit approximation error. |
| What distinguishes an operational twin? | Asset linkage, state updates, uncertainty, validation and decision purpose. |
Key points
- Start from the accountable engineering decision and its consequence.
- Compare against a transparent non-AI baseline.
- Validate on a split that represents intended use and retain human authority.
Source references recorded by the supplied chapter
- Peer-reviewed physics-informed machine-learning review listed as RES-09.
- ETH Zürich scientific machine-learning curriculum comparator, CUR-02.
- MIT and Stanford civil-engineering AI curriculum comparators.