Choose your reading route
eduKateSG · Why Mathematics? · Multi-Fidelity Reliability Modelling: Co-Kriging, Bias Correction, Budget Allocation and Validation
Start with the question, audit the worked example, or jump to the assumptions and practice. Every chapter stays open for uninterrupted reading.
Understand the question
Define the quantity and follow the mechanism.Route 2 · Chapters 6–10
Reproduce the example
Audit the numbers one step at a time.Route 3 · Chapters 11–15
Check assumptions
See where design and calibration matter.Route 4 · Chapters 16–20
Explore applications
Connect the method to a concrete investigation.Route 5 · Chapters 21–25
Study and verify
Practise, inspect failure cases and read further.
Full chapter index · Practice and checking · Mathematics Learning Hub
Why mathematics matters in reliability modelling becomes practical when evidence comes at different costs and levels of detail. A low-fidelity calculation may be cheap enough for thousands of runs, while a high-fidelity simulation or physical test is accurate in important ways but scarce. Multi-fidelity methods combine them through correlations, bias models, co-kriging and careful budget allocation.
This article shows how that combination works and why more data are not automatically better data. Low fidelity is useful only when its relationship to the high-fidelity target is measured. A large biased dataset can overwhelm a few trustworthy observations if the model fuses them carelessly.
Full chapter index
Chapters 1–5 · Understand the question
Chapters 6–10 · Reproduce the example
Chapters 11–15 · Check assumptions
Chapters 16–20 · Explore applications
Chapters 21–25 · Study and verify
Fidelity describes how a model represents relevant physics, resolution and evidence for a defined quantity and use. A coarse beam model, fine finite-element model and physical strain test form three possible levels. Fidelity is not a universal ranking across every output. A practical next step is to define fidelity relative to a quantity of interest.
Start with the mechanism: Multi-fidelity analysis exploits large cost differences while reserving expensive evaluations for calibration and checking. Consider this example: A coarse run costs one minute, a fine run two hours and a test two days. The limit is equally important: Wall-clock cost omits setup, licensing and specialist review. Students can test the idea when they record total resource cost per level.
Both models predict increasing stress with load, though the coarse model is biased low. The mathematics behind it is precise: A cheap model helps when it tracks trends or shares structure with the high-fidelity response. Yet the result has a boundary. A cheap model with weak correlation can add confusion rather than information. For a reliable investigation, estimate cross-fidelity correlation on paired points.
A common model writes high fidelity as a scaled low-fidelity process plus a discrepancy process. In an illustrative case, y_H(x)=ρy_L(x)+δ(x) separates shared trend from correction. This should not be overread because the linear relation may fail across regimes. A careful team will plot paired residuals and test alternatives. Primary research is available from https://www.sandia.gov/research/publications/details/validation-study-of-the-multi-fidelity-toolkit-2021-11-01/.
The coefficient ρ describes how the low-fidelity trend maps to the high-fidelity scale. If ρ=1.2, a one-unit low-fidelity change corresponds to 1.2 units before discrepancy. A single global ρ may hide input-dependent bias. A practical next step is to compare constant and varying scale models.
Start with the mechanism: The discrepancy δ(x) captures systematic differences not explained by scaling. Consider this example: A coarse mesh misses a stress concentration near a notch. The limit is equally important: A flexible discrepancy can absorb noise or make components non-identifiable. Students can test the idea when they use physical reasoning and validation points.
Run the coarse code at all 60 points and the fine code at 12 of those points. The mathematics behind it is precise: High-fidelity points are often placed within the low-fidelity design so levels can be compared directly. Yet the result has a boundary. Strict nesting may be infeasible for different domains or failed runs. For a reliable investigation, record pairing and missingness explicitly.
At one input the low-fidelity output is 10, ρ is 1.1 and estimated discrepancy is 2. In an illustrative case, The high-fidelity prediction is 1.1×10+2=13. This should not be overread because this point estimate omits uncertainty in ρ, discrepancy and both models. A careful team will propagate all estimated uncertainty.
Two levels can be strongly correlated while having large systematic bias. Predictions rise together but the cheap model is always 20% low. Correlation alone does not establish calibrated values. A practical next step is to report bias, scale and residual spread.
Start with the mechanism: Information-fusion estimators use correlated cheap outputs to reduce uncertainty for a fixed expensive-sample budget. Consider this example: A control-variate-style estimator adjusts the high-fidelity mean using many low-fidelity runs. The limit is equally important: Poor covariance estimates can increase variance. Students can test the idea when they estimate efficiency with repeated experiments.
A coarse model predicts safe where the fine model finds local failure. The mathematics behind it is precise: For reliability, fidelity fusion must be accurate near the limit-state boundary and in the tail. Yet the result has a boundary. Good average response agreement can hide wrong tail classification. For a reliable investigation, compare failure indicators on paired cases.
Recent research combines multi-fidelity information fusion with importance sampling or splitting for rare failures. In an illustrative case, High-fidelity runs concentrate in an important failure-prone region while cheap runs guide proposals. This should not be overread because biased proposals or wrong corrections distort the estimate. A careful team will publish weights and effective sample size. A second primary or official research source is https://doi.org/10.1016/j.probengmech.2022.103291.
Hierarchical kriging uses a low-fidelity surrogate as trend information for a high-fidelity GP. Sandia reports hierarchical kriging within a validated multi-fidelity aerodynamics toolkit. Toolkit validation for one flow problem does not certify another reliability model. A practical next step is to repeat verification and validation for the new use.
Start with the mechanism: Tests may anchor simulation levels but introduce measurement error and laboratory variability. Consider this example: Wind-tunnel measurements check aerodynamic predictions with validation uncertainty. The limit is equally important: Experiments are not perfect truth and may cover a narrow regime. Students can test the idea when they model observation error separately.
Mesh convergence supports numerical verification, while held-out tests support validation. The mathematics behind it is precise: Verification checks code and numerical solution; validation compares model outputs with relevant physical evidence. Yet the result has a boundary. Agreement at calibrated points is not independent validation. For a reliable investigation, reserve validation cases before fitting.
Each level adds input uncertainty, numerical error, parameter uncertainty, discrepancy and observation uncertainty. In an illustrative case, A final interval combines several named components rather than one unexplained width. This should not be overread because double counting or omitting covariance can distort intervals. A careful team will draw an uncertainty budget.
Choose counts at each fidelity to reduce decision uncertainty per unit cost. Spend 120 minutes on 60 coarse runs or one fine run depending on marginal value. Optimal allocations depend on unknown correlations and costs. A practical next step is to pilot first and update the plan.
Start with the mechanism: At each candidate, decide both where to sample and which fidelity to use. Consider this example: Use cheap runs for broad coverage and fine runs where discrepancy or failure classification is uncertain. The limit is equally important: A rule may repeatedly choose the cheap level and never correct bias. Students can test the idea when they guarantee high-fidelity checks and stopping criteria.
Ten fine runs alone cost 1,200 minutes, so the mixed plan saves 500 minutes before setup costs. The mathematics behind it is precise: Fifty coarse runs at 2 minutes and five fine runs at 120 minutes cost 700 minutes. Yet the result has a boundary. Lower cost does not prove equal accuracy. For a reliable investigation, compare held-out error and decision quality.
Create a simple cheap line and an expensive curved function, then use paired points to estimate scale and correction. In an illustrative case, Change the number of expensive points and watch uncertainty near a threshold. This should not be overread because a toy function lacks numerical and experimental error. A careful team will explain which level changed the decision.
Multi-fidelity modelling develops judgement about evidence quality, ratios, functions and budgets. Ask what each level simplifies and whether the simplification matters to the decision. More rows from a weak model do not become stronger evidence by repetition. A practical next step is to begin with two measuring instruments of different precision.
Start with the mechanism: High fidelity is not perfect truth, low fidelity is not useless and fusion is not simple averaging. Consider this example: A detailed simulation can still omit the dominant mechanism. The limit is equally important: Visual detail is not validation. Students can test the idea when they require paired comparisons and uncertainty.
A cheap model reverses trend exactly where failure occurs. The mathematics behind it is precise: If levels have weak or unstable relationships, analyse them separately or improve the physical model. Yet the result has a boundary. Forcing fusion can create false confidence. For a reliable investigation, use a diagnostic gate before combination.
Sandia documents formal validation of a multi-fidelity toolkit, while primary reliability research develops information-fusion estimators. In an illustrative case, Read both sources with the surrogate-model and uncertainty-quantification owners. This should not be overread because a method paper does not set the right fidelity hierarchy for every system. A careful team will use the hub to review correlation and weighted means. Continue with the Mathematics Learning Hub and the previous reliability article.
FAQs should cover level definitions, paired designs, co-kriging, discrepancy, cost, failure tails and validation. Answers should identify the high-fidelity target. Do not call a cheap model accurate merely because it is correlated. A practical next step is to publish the fidelity map and budget.
Extended practice laboratory
This laboratory turns multi-fidelity reliability modelling into a sequence of checkable decisions. Students should keep a ledger of definitions, calculations, assumptions and later evidence. A parent or educator can ask what the model knew at the time and what would make the conclusion change. The goal is not to imitate advanced software; it is to understand the mathematical mechanism well enough to detect a wrong denominator, a leaked timestamp, an unsupported extrapolation or an overclaim.
A strong project on multi-fidelity reliability modelling includes a simple baseline, one worked example, a sensitivity check and an untouched evaluation. It reports limits with the same care as results. This is how school algebra, probability, graphs and functions become transferable reasoning tools.
For a capstone, build a small reproducible investigation of multi-fidelity reliability modelling. Freeze the question and data contract first. Preserve every intermediate value so another student can recompute the result. Compare at least two defensible methods, evaluate on later or held-out evidence, and include a failure case. End with a one-page decision record: what is supported, what remains uncertain, what action is allowed, and what new evidence would trigger review. The capstone succeeds when its reasoning survives a changed example, not when it produces the most impressive chart. A multi-fidelity capstone should define every level by equations, resolution, cost and intended output. Run paired low- and high-fidelity cases across ordinary and failure-prone regimes, estimate scale and discrepancy, and inspect whether their relationship changes across the domain. Compare direct high-fidelity estimation, simple correction, co-kriging and a control-variate-style estimator under the same total budget. Reserve physical or independent high-fidelity cases for validation. Report bias, correlation, variance reduction, failure-sign disagreement and total cost separately. Stress-test the fusion after weakening cross-fidelity correlation and increasing observation error. Draw an uncertainty budget that prevents numerical error, model discrepancy and measurement uncertainty from being double counted. The final fidelity map should identify where cheap evidence helps, where it merely repeats bias, when a high-fidelity run is mandatory and when the levels must remain separate.
Verification and validation deeper audit
Use verification and validation to practise sensitivity analysis. Verification checks code and numerical solution; validation compares model outputs with relevant physical evidence. Start from this example: Mesh convergence supports numerical verification, while held-out tests support validation. Change one assumption while keeping the others fixed and predict the direction before calculating. Agreement at calibrated points is not independent validation. Reserve validation cases before fitting. Report whether the decision is stable or depends heavily on the changed assumption. Record the result in plain language for a student and technical language for a reviewer. The two versions should agree on the quantity and limitation even though their vocabulary differs.
Uncertainty bookkeeping deeper audit
Build a small counterexample for uncertainty bookkeeping. Each level adds input uncertainty, numerical error, parameter uncertainty, discrepancy and observation uncertainty. A final interval combines several named components rather than one unexplained width. Then create a nearby case where the shortcut fails. This matters because double counting or omitting covariance can distort intervals. Draw an uncertainty budget. Explain the failure in plain language and connect it to the exact mathematical step that changed. Record the result in plain language for a student and technical language for a reviewer. The two versions should agree on the quantity and limitation even though their vocabulary differs.
Budget allocation deeper audit
Treat budget allocation as a model-checking problem. Choose counts at each fidelity to reduce decision uncertainty per unit cost. Spend 120 minutes on 60 coarse runs or one fine run depending on marginal value. Preserve inputs, intermediate values and output. Optimal allocations depend on unknown correlations and costs. Pilot first and update the plan. Evaluate the same rule on later or held-out evidence and separate numerical correctness from real-world validity. Record the result in plain language for a student and technical language for a reviewer. The two versions should agree on the quantity and limitation even though their vocabulary differs.
Adaptive level selection deeper audit
Connect the mathematics of adaptive level selection to a decision. At each candidate, decide both where to sample and which fidelity to use. Use cheap runs for broad coverage and fine runs where discrepancy or failure classification is uncertain. Calculate the quantity, then translate it into an allowed action and an action that the evidence does not support. A rule may repeatedly choose the cheap level and never correct bias. Guarantee high-fidelity checks and stopping criteria. This distinction prevents an accurate calculation from becoming an overclaim. Record the result in plain language for a student and technical language for a reviewer. The two versions should agree on the quantity and limitation even though their vocabulary differs.
Worked cost comparison deeper audit
Turn worked cost comparison into a transparent worksheet. Fifty coarse runs at 2 minutes and five fine runs at 120 minutes cost 700 minutes. Ten fine runs alone cost 1,200 minutes, so the mixed plan saves 500 minutes before setup costs. List every quantity, unit and timestamp, then recompute the example after changing one input. Lower cost does not prove equal accuracy. Compare held-out error and decision quality. Finish by naming the output as description, prediction, diagnosis or causal evidence and justify that label. Record the result in plain language for a student and technical language for a reviewer. The two versions should agree on the quantity and limitation even though their vocabulary differs.
Student activity deeper audit
Begin with the operational question behind student activity. Create a simple cheap line and an expensive curved function, then use paired points to estimate scale and correction. Now examine the illustrative case: Change the number of expensive points and watch uncertainty near a threshold. Draw the relevant timeline or graph before calculating. The important boundary is that a toy function lacks numerical and experimental error. Explain which level changed the decision. Compare a simple baseline and record what evidence would reverse the conclusion. Record the result in plain language for a student and technical language for a reviewer. The two versions should agree on the quantity and limitation even though their vocabulary differs.
Parent and educator guidance deeper audit
For this case, audit the mechanism rather than memorising vocabulary. Multi-fidelity modelling develops judgement about evidence quality, ratios, functions and budgets. Ask what each level simplifies and whether the simplification matters to the decision. Ask which denominator, risk set, neighbourhood or candidate pool makes the result meaningful. More rows from a weak model do not become stronger evidence by repetition. Begin with two measuring instruments of different precision. A second student should be able to reproduce the answer from the saved ledger alone. Record the result in plain language for a student and technical language for a reviewer. The two versions should agree on the quantity and limitation even though their vocabulary differs.
Practice case 1: What fidelity means
Setup: Use a small example for what fidelity means.
Task: Define fidelity relative to a quantity of interest.
Checking target: Fidelity is not a universal ranking across every output.
Turn what fidelity means into a transparent worksheet. Fidelity describes how a model represents relevant physics, resolution and evidence for a defined quantity and use. A coarse beam model, fine finite-element model and physical strain test form three possible levels. List every quantity, unit and timestamp, then recompute the example after changing one input. Fidelity is not a universal ranking across every output. Define fidelity relative to a quantity of interest. Finish by naming the output as description, prediction, diagnosis or causal evidence and justify that label.
Practice case 2: Cost hierarchy
Setup: Use a small example for cost hierarchy.
Task: Record total resource cost per level.
Checking target: Wall-clock cost omits setup, licensing and specialist review.
Begin with the operational question behind cost hierarchy. Multi-fidelity analysis exploits large cost differences while reserving expensive evaluations for calibration and checking. Now examine the illustrative case: A coarse run costs one minute, a fine run two hours and a test two days. Draw the relevant timeline or graph before calculating. The important boundary is that wall-clock cost omits setup, licensing and specialist review. Record total resource cost per level. Compare a simple baseline and record what evidence would reverse the conclusion.
Practice case 3: Low-fidelity usefulness
Setup: Use a small example for low-fidelity usefulness.
Task: Estimate cross-fidelity correlation on paired points.
Checking target: A cheap model with weak correlation can add confusion rather than information.
For this case, audit the mechanism rather than memorising vocabulary. A cheap model helps when it tracks trends or shares structure with the high-fidelity response. Both models predict increasing stress with load, though the coarse model is biased low. Ask which denominator, risk set, neighbourhood or candidate pool makes the result meaningful. A cheap model with weak correlation can add confusion rather than information. Estimate cross-fidelity correlation on paired points. A second student should be able to reproduce the answer from the saved ledger alone.
Practice case 4: Autoregressive co-kriging
Setup: Use a small example for autoregressive co-kriging.
Task: Plot paired residuals and test alternatives.
Checking target: The linear relation may fail across regimes.
Use autoregressive co-kriging to practise sensitivity analysis. A common model writes high fidelity as a scaled low-fidelity process plus a discrepancy process. Start from this example: y_H(x)=ρy_L(x)+δ(x) separates shared trend from correction. Change one assumption while keeping the others fixed and predict the direction before calculating. The linear relation may fail across regimes. Plot paired residuals and test alternatives. Report whether the decision is stable or depends heavily on the changed assumption.
Practice case 5: Scaling parameter
Setup: Use a small example for scaling parameter.
Task: Compare constant and varying scale models.
Checking target: A single global ρ may hide input-dependent bias.
Build a small counterexample for scaling parameter. The coefficient ρ describes how the low-fidelity trend maps to the high-fidelity scale. If ρ=1.2, a one-unit low-fidelity change corresponds to 1.2 units before discrepancy. Then create a nearby case where the shortcut fails. This matters because a single global ρ may hide input-dependent bias. Compare constant and varying scale models. Explain the failure in plain language and connect it to the exact mathematical step that changed.
Practice case 6: Discrepancy function
Setup: Use a small example for discrepancy function.
Task: Use physical reasoning and validation points.
Checking target: A flexible discrepancy can absorb noise or make components non-identifiable.
Treat discrepancy function as a model-checking problem. The discrepancy δ(x) captures systematic differences not explained by scaling. A coarse mesh misses a stress concentration near a notch. Preserve inputs, intermediate values and output. A flexible discrepancy can absorb noise or make components non-identifiable. Use physical reasoning and validation points. Evaluate the same rule on later or held-out evidence and separate numerical correctness from real-world validity.
Practice case 7: Nested designs
Setup: Use a small example for nested designs.
Task: Record pairing and missingness explicitly.
Checking target: Strict nesting may be infeasible for different domains or failed runs.
Connect the mathematics of nested designs to a decision. High-fidelity points are often placed within the low-fidelity design so levels can be compared directly. Run the coarse code at all 60 points and the fine code at 12 of those points. Calculate the quantity, then translate it into an allowed action and an action that the evidence does not support. Strict nesting may be infeasible for different domains or failed runs. Record pairing and missingness explicitly. This distinction prevents an accurate calculation from becoming an overclaim.
Practice case 8: A correction example
Setup: Use a small example for a correction example.
Task: Propagate all estimated uncertainty.
Checking target: This point estimate omits uncertainty in ρ, discrepancy and both models.
Turn a correction example into a transparent worksheet. At one input the low-fidelity output is 10, ρ is 1.1 and estimated discrepancy is 2. The high-fidelity prediction is 1.1×10+2=13. List every quantity, unit and timestamp, then recompute the example after changing one input. This point estimate omits uncertainty in ρ, discrepancy and both models. Propagate all estimated uncertainty. Finish by naming the output as description, prediction, diagnosis or causal evidence and justify that label.
Practice case 9: Correlation is not agreement
Setup: Use a small example for correlation is not agreement.
Task: Report bias, scale and residual spread.
Checking target: Correlation alone does not establish calibrated values.
Begin with the operational question behind correlation is not agreement. Two levels can be strongly correlated while having large systematic bias. Now examine the illustrative case: Predictions rise together but the cheap model is always 20% low. Draw the relevant timeline or graph before calculating. The important boundary is that correlation alone does not establish calibrated values. Report bias, scale and residual spread. Compare a simple baseline and record what evidence would reverse the conclusion.
Practice case 10: Variance reduction
Setup: Use a small example for variance reduction.
Task: Estimate efficiency with repeated experiments.
Checking target: Poor covariance estimates can increase variance.
For this case, audit the mechanism rather than memorising vocabulary. Information-fusion estimators use correlated cheap outputs to reduce uncertainty for a fixed expensive-sample budget. A control-variate-style estimator adjusts the high-fidelity mean using many low-fidelity runs. Ask which denominator, risk set, neighbourhood or candidate pool makes the result meaningful. Poor covariance estimates can increase variance. Estimate efficiency with repeated experiments. A second student should be able to reproduce the answer from the saved ledger alone.
Practice case 11: Failure-probability target
Setup: Use a small example for failure-probability target.
Task: Compare failure indicators on paired cases.
Checking target: Good average response agreement can hide wrong tail classification.
Use failure-probability target to practise sensitivity analysis. For reliability, fidelity fusion must be accurate near the limit-state boundary and in the tail. Start from this example: A coarse model predicts safe where the fine model finds local failure. Change one assumption while keeping the others fixed and predict the direction before calculating. Good average response agreement can hide wrong tail classification. Compare failure indicators on paired cases. Report whether the decision is stable or depends heavily on the changed assumption.
Practice case 12: Importance sampling fusion
Setup: Use a small example for importance sampling fusion.
Task: Publish weights and effective sample size.
Checking target: Biased proposals or wrong corrections distort the estimate.
Build a small counterexample for importance sampling fusion. Recent research combines multi-fidelity information fusion with importance sampling or splitting for rare failures. High-fidelity runs concentrate in an important failure-prone region while cheap runs guide proposals. Then create a nearby case where the shortcut fails. This matters because biased proposals or wrong corrections distort the estimate. Publish weights and effective sample size. Explain the failure in plain language and connect it to the exact mathematical step that changed.
Practice case 13: Hierarchical kriging
Setup: Use a small example for hierarchical kriging.
Task: Repeat verification and validation for the new use.
Checking target: Toolkit validation for one flow problem does not certify another reliability model.
Treat hierarchical kriging as a model-checking problem. Hierarchical kriging uses a low-fidelity surrogate as trend information for a high-fidelity GP. Sandia reports hierarchical kriging within a validated multi-fidelity aerodynamics toolkit. Preserve inputs, intermediate values and output. Toolkit validation for one flow problem does not certify another reliability model. Repeat verification and validation for the new use. Evaluate the same rule on later or held-out evidence and separate numerical correctness from real-world validity.
Practice case 14: Physical experiments as a level
Setup: Use a small example for physical experiments as a level.
Task: Model observation error separately.
Checking target: Experiments are not perfect truth and may cover a narrow regime.
Connect the mathematics of physical experiments as a level to a decision. Tests may anchor simulation levels but introduce measurement error and laboratory variability. Wind-tunnel measurements check aerodynamic predictions with validation uncertainty. Calculate the quantity, then translate it into an allowed action and an action that the evidence does not support. Experiments are not perfect truth and may cover a narrow regime. Model observation error separately. This distinction prevents an accurate calculation from becoming an overclaim.