Engineering & Technology · Specialist Engineering & Emerging Technologies

Data-assisted design, AI integration and trustworthy digital models

A design tool proposes a promising component and a digital model displays it convincingly. The result still needs a physical question, a valid representation and evidence for its intended use. Data and computation can support engineering judgement while introducing their own sources of error.

Revisit material structure, heat and energy balances, electronics, uncertainty and system integration. Numerical examples are invented teaching cases, not nuclear, manufacturing or device operating specifications.

Back to the chapter

After reading, you can

Use data-assisted predictions within their validated context and preserve the complete digital evidence chain.

Separate the model’s roles

A numerical model may express physical balances; a data-assisted model may learn a relationship from examples; a hybrid combines elements of both. Each needs a defined input, output and intended decision. An optimizer searches the supplied representation, so an attractive optimum can exploit its missing constraints or inaccurate regions.

A digital twin links a digital representation to a particular physical system and its relevant data and lifecycle context. A detailed visual model alone does not establish that link or its credibility. Record what is represented, how it is updated and what decisions its predictions are intended to support.

Evaluate independence and the range of use

Data used to select a model or adjust its parameters are not an independent evaluation of the final selection. Splits also need to respect identities and time: nearby observations from the same asset may share information. The intended future use determines which independent assets, periods and conditions must be represented.

Prediction error within a familiar range need not describe an unseen condition. Changes in material, operating regime, sensors or data definitions can invalidate the earlier relationship. Report errors with their units, denominators and case coverage rather than treating a single average as proof of broad reliability.

Verification asks whether the chosen calculation was implemented correctly. Validation asks whether it is adequate for the intended real-world question. A machine-learning label or a plausible simulation cannot settle either question by itself. The original graph retains the local comparisons and a separate new-range case.

Horizontal reference 0–70 arbitrary units at 7 SVG units/unit, vertical prediction 0–70 at 3.5 units/unit; different drawing scales, same numeric ranges. Dashed y=x agreement line is not 45 degrees because plot aspect differs. Circles at (10, 9), (20, 21), (30, 31), (40, 39) have residuals −1, +1, +1, −1 and MAE=RMSE=1. Square at (60, 48) has residual −12, relative error −20%. These are supplied original outputs and independent references, not an executed AI model or statistical generalization study.

Integrate a trustworthy decision process

Keep units, data provenance, processing and model versions linked to the physical configuration. Missing or corrupted inputs, stale records, access controls and unintended model changes can affect the decision. A useful system also needs monitoring, a response to uncertain outputs and responsibility for changes.

The NIST 2023 framework describes trustworthiness as context-dependent and broader than accuracy alone. Explainability, privacy, security, fairness and human oversight may create additional requirements or tradeoffs. Their relevance depends on the application; a model’s average error cannot certify the whole decision process.

Check local error and one changed condition

Supply independent reference responses 10, 20, 30 and 40 arbitrary units, with predictor outputs 9, 21, 31 and 39. Define residual as prediction minus reference. The residuals are −1, +1, +1, −1, so mean residual is zero, mean absolute error 1 and root-mean-square error 1. The zero mean does not mean every prediction is exact.

For a separate new-range case, supply reference 60 and prediction 48. Residual is −12; relative error is −12/60=−20%. The first four small errors did not bound this fifth error. No confidence limit or population distribution has been established by five constructed values.

Suppose a design requires a property of at least 40 units. A predicted 42 with an independently established 35 would fail that requirement despite the favourable prediction. This is a separate illustration of a decision consequence, not a tolerance or a result from an executed model.

Try the reasoning

Does zero mean residual prove accurate individual predictions?
Show the answer and reasoning
No. Positive and negative errors can cancel; their sizes and distribution still matter.
Why can a random observation split be misleading?
Show the answer and reasoning
Training and evaluation observations may share an asset, time context or other information that makes them dependent for the intended use.
Does an optimizer validate the model it searches?
Show the answer and reasoning
No. It finds favourable values within the supplied representation, including any errors or missing constraints.

What to carry forward

Connect digital outputs to independent evidence, a defined context and accountable decisions.

References

Teaching resources and research supporting this lesson’s explanations and examples.