Simulation-Informed Decisions

Simulation as a Decision Tool, Not an End in Itself

Simulation in engineering is not new. What is new — and what digital engineering demands — is using simulation systematically within a framework where simulations are connected to the system model, configured from model data, and their results feed back into traceable engineering decisions.

The danger is treating simulation as a standalone activity: an analyst receives a request, runs a model, produces a report, and hands it off. In that workflow, the simulation is disconnected from the decision it should inform. The analyst does not know what decision hinges on the result. The decision-maker does not know what assumptions the simulation made. The report sits in a folder, referenced vaguely ("analysis confirms the design is adequate") rather than precisely ("simulation predicts a thermal margin of 8 degrees C at the worst-case duty cycle, with a model uncertainty of plus or minus 3 degrees C").

The Simulation Pyramid

Not all simulations serve the same purpose. Using the wrong level of fidelity for the decision at hand wastes resources (too much fidelity) or produces unreliable results (too little fidelity). The simulation pyramid provides a framework for matching fidelity to decision need.

Level 1: Screening

Fidelity: Low. Analytical equations, parametric models, lookup tables, linear approximations.

Speed: Milliseconds to seconds.

Purpose: Eliminate obviously poor alternatives quickly. When you have twenty candidate configurations and need to narrow to five, screening-level analysis separates the viable from the non-viable without investing in detailed modeling.

Example: Mass and power budget calculations from parametric models in the system model. If a configuration exceeds the mass budget by 40%, no detailed structural analysis is needed to eliminate it.

Connection to MBSE: Screening analyses often run directly within the system model's parametric framework. The model-defined properties (component masses, power draws, cost estimates) feed the screening calculations without any tool integration — the model does the work.

Level 2: Exploration

Fidelity: Medium. Reduced-order models, simplified finite element analysis, 1D thermal networks, lumped-parameter dynamics.

Speed: Minutes to hours.

Purpose: Compare design alternatives and identify sensitivities. This is the trade study workhorse — fidelity is sufficient to differentiate alternatives, fast enough to evaluate multiple configurations, and structured enough to feed quantitative criteria.

Example: Simplified thermal analysis of five heat sink geometries to compare steady-state temperatures at the critical component. The models are approximate but consistent — the same simplifications apply to all alternatives, making relative comparison valid even if absolute accuracy is limited.

Connection to MBSE: Exploration models read their inputs (geometry parameters, material properties, boundary conditions) from the system model via API or export. Results write back to the model as performance predictions, feeding trade study criteria.

Level 3: Validation

Fidelity: High. Full finite element analysis, computational fluid dynamics, multi-physics coupling, Monte Carlo methods.

Speed: Hours to days.

Purpose: Validate the selected design against requirements. After the trade study narrows to a preferred alternative, high-fidelity simulation confirms that the design meets its performance obligations with appropriate margin.

Example: Full conjugate heat transfer CFD of the selected heat sink design under the complete set of thermal load cases, including transient duty cycles and worst-case ambient conditions.

Connection to MBSE: Validation simulations are formally linked to requirements in the model. Each simulation run traces to the requirement it verifies, the model version it used as input, and the pass/fail verdict it produced. This traceability is what distinguishes validation from exploration — the result is not informational, it is a verification record.

Level 4: Certification

Fidelity: Highest. Calibrated and validated models with documented uncertainty quantification. May require correlation with physical test data.

Speed: Days to weeks.

Purpose: Satisfy regulatory or contractual requirements. Some domains (aerospace, automotive safety, nuclear) allow or require simulation-based evidence for certification. The simulation must demonstrate not just that the design passes but that the model is credible — its assumptions are documented, its predictions have been compared to test data, and its uncertainty is bounded.

Example: Crash simulation for automotive safety certification per FMVSS 208, with documented model validation against sled test data and uncertainty analysis on occupant injury metrics.

Connection to MBSE: Certification models are the most tightly governed. Model configurations, solver settings, mesh parameters, and validation evidence are all part of the verification record linked to the system model. The digital thread for a certification simulation must be unbroken from requirement through analysis to evidence.

Connecting Simulations to the System Model

In a digital engineering workflow, simulations are not standalone activities. They are nodes in the digital thread, with defined inputs and outputs that connect to the system model.

Inputs from the model:

• Geometry from the CAD model (driven by system model parameters)
• Material properties from the system model's component definitions
• Load cases derived from the requirements model
• Interface conditions from the architecture model (boundary conditions at component interfaces)

Outputs to the model:

• Performance predictions (temperature, stress, deflection, efficiency)
• Margin calculations (how close to the requirement threshold?)
• Pass/fail verdicts against specific requirements
• Sensitivity data (which parameters drive the result most strongly?)

Traceability: Each simulation run should be traceable to the model version it used as input, the tool and configuration it ran with, the results it produced, and the decision it informed. Without this traceability, a simulation result is an orphan — a number without provenance.

Surrogate Models for Rapid Evaluation

High-fidelity simulations are accurate but slow. When a decision requires evaluating hundreds or thousands of configurations — as in optimization or uncertainty quantification — running the full-fidelity model for each configuration is impractical.

Surrogate models bridge this gap. The approach:

1. Run a set of high-fidelity simulations across the design space using a design of experiments (DoE) sampling strategy
2. Train a mathematical approximation (Gaussian process, polynomial response surface, neural network) on the simulation inputs and outputs
3. Use the surrogate to predict outputs for new configurations in milliseconds

Surrogate models are predictive tools, not replacements for high-fidelity analysis. They are accurate within the training region and unreliable outside it. They are useful for exploration and optimization but should not be used for validation or certification without verification against the full-fidelity model.

This connects directly to DGTLENG 106 (Automation and Digital Pipelines), where surrogate model training can be automated as part of a simulation pipeline — run the DoE, train the surrogate, validate its accuracy, and deploy it for rapid evaluation.

The Traps

Analysis Paralysis

Running more and more simulations without making a decision. The team requests higher fidelity, additional load cases, more mesh refinement — not because the decision requires it but because deciding is uncomfortable and analysis is familiar.

The counter: Define decision criteria before running the simulation. Specify what result would confirm the alternative is acceptable and what result would eliminate it. When the simulation delivers that result, decide.

False Precision

A simulation reports stress to six decimal places. That does not mean the result is accurate to six decimal places. Every simulation carries uncertainty from modeling assumptions, discretization, input variability, and solver tolerances. Reporting "the maximum stress is 247.382914 MPa" when the model uncertainty is plus or minus 15% is misleading.

The counter: Always report simulation results with their uncertainty. "Maximum stress is approximately 247 MPa with a model uncertainty of plus or minus 37 MPa based on mesh convergence and material property variability." The precision of the output should match the confidence in the input.

Validation Theater

Running simulations to confirm a decision already made, rather than to inform a decision not yet made. The result is predetermined — if the simulation disagrees with the preferred outcome, the analyst is told to "check the model" until it agrees.

The counter: The value of simulation is in surprise — results that challenge assumptions. A simulation that confirms what you already believed has lower information value than one that reveals something unexpected. Treat disagreement between simulation and expectation as a learning opportunity, not a problem to be fixed.

Assessment