Part I — Simulation-Driven Engineering: From Physics to Intelligent Optimisation

Introduction

Every engineering design process is fundamentally a search problem.

Engineers seek combinations of geometry, materials, and operating conditions that maximise performance while satisfying physical and practical constraints. As products become increasingly complex, the number of feasible designs grows far beyond what can be explored manually, making computational simulation an essential component of modern engineering.

However, the increasing capability of simulation tools has introduced a different challenge.

The limiting factor is no longer whether a design can be analysed, but whether the design space can be explored efficiently.

Figure 1. The modern simulation-driven engineering workflow. Computational optimisation forms an iterative loop between parameterised models, simulation results, and validation.

A conventional engineering workflow typically evaluates a small number of candidate designs through an iterative process. This approach relies heavily on engineering intuition to determine which designs should be investigated.

It is highly effective for incremental optimisation, where previous experience provides a reasonable estimate of how design changes will influence performance.

However, many modern engineering problems no longer exist within small, intuitive design spaces.

Advanced systems frequently contain hundreds of interacting variables spanning:

• Geometry
• Material properties
• Operating conditions
• Control parameters
• Environmental constraints

The relationship between these variables and system performance is often:

• Nonlinear
• Strongly coupled
• Difficult to predict without numerical evaluation

The engineering problem therefore changes from:

Determining whether a design works

to:

Determining which designs are worth evaluating.

---

Engineering as a Search Problem

To formalise this concept, the engineering design process can be represented as a mathematical mapping.

A design is defined by a vector of $n$ design variables:

$$\mathbf{x} = \begin{bmatrix} x_1\\ x_2\\ \vdots\\ x_n \end{bmatrix}$$

These variables represent the physical parameters under the engineer's control, such as:

• Geometrical dimensions
• Material properties
• Operating conditions

When this design is evaluated by a physical model, it produces a vector of $m$ performance metrics:

$$\mathbf{y} = \begin{bmatrix} y_1\\ y_2\\ \vdots\\ y_m \end{bmatrix}$$

where outputs may include:

• Efficiency
• Mass
• Temperature
• Stress
• Vibration
• Cost

Figure 2. The engineering search problem formulated as a mathematical mapping.

This relationship is expressed as:

$$\mathbf{y}=f(\mathbf{x})$$

For simple analytical systems, $f$ may be evaluated directly.

However, for complex engineering systems, this function is replaced by computational models involving thousands or millions of coupled equations.

The cost of evaluating $f(\mathbf{x})$ therefore becomes the dominant limitation within the optimisation process.

This creates a fundamental conflict:

The challenge is therefore not simply developing more accurate simulations.

It is determining how different levels of simulation fidelity should be combined throughout the design process.

---

Physics-Based Simulation

Before an optimisation algorithm can explore a design space, the physical system must first be represented mathematically.

The quality of any engineering optimisation is fundamentally limited by the fidelity of the underlying model.

An optimisation algorithm, regardless of sophistication, can only identify the optimum of the model presented to it.

If important physical phenomena are omitted, the resulting design may be mathematically optimal while performing poorly in reality.

Regardless of system complexity, numerical simulations originate from several fundamental conservation laws:

• Conservation of mass
• Conservation of momentum
• Conservation of energy
• Conservation of charge

At the lowest level of representation, these become ordinary differential equations (ODEs) describing lumped system behaviour.

---

Conservation Laws

For a general control volume, conservation of mass is:

$$\frac{dm}{dt} = \dot{m}_{in} - \dot{m}_{out}$$

The conservation of energy is:

$$\frac{dU}{dt} = \dot{Q} - P\frac{dV}{dt} + \sum_i \dot{m}_i h_i$$

where:

• $U$ is internal energy
• $\dot{Q}$ is heat transfer rate
• $P$ is pressure
• $V$ is volume
• $\dot{m}_i$ is mass flow rate
• $h_i$ is specific enthalpy

As the engineering question becomes more specific—requiring knowledge of:

• Boundary layers
• Turbulent mixing
• Localised thermal stresses

lumped assumptions become insufficient.

The system transitions from ODE-based representations into spatially distributed partial differential equations (PDEs).

---

Governing PDE Models

For fluid systems, the governing equations become the Navier–Stokes equations:

$$\frac{\partial U}{\partial t} + \nabla\cdot F_c = \nabla\cdot F_v + S$$

where $U$ contains:

• Density,
• Momentum, and
• Total Energy

Note: This is a compact and consercative form of the compressible Navier-Stokes equations.

For thermal systems:

$$\rho c_p \frac{\partial T}{\partial t} = \nabla\cdot(k\nabla T) + Q$$

For structural systems:

$$\nabla\cdot\sigma + f = \rho \frac{\partial^2 u}{\partial t^2}$$

Modern engineering systems are rarely governed by a single physical domain.

The output of one physical model becomes the input of another:

Combustion
      ↓
Temperature Field
      ↓
Thermal Expansion
      ↓
Structural Stress
      ↓
Fluid Interaction

A complete multiphysics simulation therefore assembles these models into a coupled nonlinear system where each physical process influences the evolution of the others.

Because no single model is universally appropriate, engineers must select from a spectrum of representations.

---

The Hierarchy of Engineering Models

Engineering models exist along a continuous spectrum of abstraction.

Increasing model fidelity introduces additional physical information, but also increases computational cost.

The objective of engineering simulation is therefore not to maximise fidelity at all times.

Instead, the objective is:

Select the model that provides sufficient information for the decision being made.

Figure 3. The hierarchy of engineering models. Increasing fidelity provides additional physical information but requires increasing computational resources.

Although this hierarchy appears linear, it should not be interpreted as a replacement sequence.

Higher-fidelity models do not eliminate the need for simpler models.

Each model answers different engineering questions.

Model selection is therefore an information problem rather than purely an accuracy problem.

---

Case Study: The Pulse Jet

To demonstrate how this hierarchy operates in practice, consider the design of a pulse jet engine.

Although mechanically simple compared with modern propulsion systems, the pulse jet represents a highly coupled multiphysics system.

A change to exhaust geometry modifies:

• Acoustic resonance
• Pressure oscillation behaviour
• Flow losses
• Thrust generation

Increasing combustion energy improves propulsion performance but simultaneously increases:

• Thermal loading
• Material constraints
• Structural requirements

The modelling hierarchy can therefore be mapped directly onto this system.

Figure 4. The modelling hierarchy applied to the pulse jet case study.

---

Analytical Model

At the lowest level, acoustic resonance frequency can be estimated algebraically:

$$f_n = \frac{(2n-1)c}{4L_t}$$

where:

• $f_n$ is the resonant frequency
• $n$ is the acoustic mode number
• $c$ is the speed of sound
• $L_t$ is the effective tube length

This provides immediate intuition regarding how exhaust length $L_t$ affects operation.

However, it ignores:

• Fluid dynamics
• Combustion behaviour
• Pressure wave interactions
• Thermal effects

---

0D Thermodynamic Model

The engine is treated as a single uniform control volume.

Spatial variation is neglected:

$$P(x,y,z,t)\approx P(t)$$

The conservation equations are applied directly to estimate:

• Energy conversion efficiency
• Pressure rise
• Operating temperature

These calculations execute in milliseconds but cannot predict:

• Velocity distributions
• Pressure wave propagation
• Localised thermal effects

---

1D Compressible Flow Model

The engine is represented as a sequence of connected sections.

Physical quantities vary along the engine axis:

$$P=P(x,t)$$

This introduces quasi-one-dimensional conservation equations capable of capturing:

• Pressure wave propagation
• Exhaust acceleration
• Flow choking

It provides the transition between thermodynamic feasibility and spatial fluid dynamics.

---

3D Reacting CFD Model

To capture:

• Turbulent mixing
• Flame structure
• Boundary-layer losses
• Local thermal stresses

the system transitions into full Navier–Stokes modelling.

The computational domain may contain millions of elements, providing an authoritative representation of a single design.

However, the simulation may require hours or days.

---

Experiment

Physical testing remains the highest physical authority for the tested configuration, but carries:

• The highest cost
• The longest lead time
• Additional manufacturing constraints

This modelling hierarchy applies universally.

Examples include:

The pulse jet simply provides a compact example of a broader engineering reality.

---

The Computational Bottleneck

The hierarchy of models provides a framework for understanding physical systems.

However, it also exposes the central limitation of modern engineering design:

Physical fidelity is computationally expensive.

Suppose a single high-fidelity multiphysics simulation of the pulse jet requires four hours to execute.

If an optimisation algorithm requires 100 simulations to converge:

$$100\times4\text{ hours} = 400\text{ hours} \approx16.7\text{ days}$$

However, if a global search requires 100,000 simulations to adequately explore the design space:

$$100000\times4\text{ hours} = 400000\text{ hours} \approx45.6\text{ years}$$

The computational scaling problem can be visualised across the modelling hierarchy.

Figure 5. Approximate logarithmic scaling of computational cost across engineering model fidelity.

---

The problem is compounded by the size of modern engineering design spaces.

Consider only seven design variables for the pulse jet.

If each variable is discretised into fifty possible values:

$$50^7 = 7.81\times10^{11}$$

possible designs exist.

Exhaustively evaluating every configuration is impossible. Even coarse discretisation rapidly becomes infeasible as dimensionality increases.

This drastic growth in possible designs is commonly referred to as the curse of dimensionality. As the number of design variables increases, the number of required evaluations grows exponentially, causing conventional search strategies to become computationally infeasible.

---

Multi-Objective Engineering Design

Engineering systems rarely possess a single optimal solution.

Increasing combustion energy may increase thrust, but simultaneously increases:

• Wall temperature
• Thermal stress
• Cooling requirements
• Material constraints

The result is therefore not a single optimum, but a Pareto frontier:

A set of solutions where improving one objective necessarily degrades another.

For a multi-objective engineering problem:

$$\min_{\mathbf{x}} \begin{bmatrix} f_1(\mathbf{x})\\ f_2(\mathbf{x})\\ \vdots\\ f_m(\mathbf{x}) \end{bmatrix}$$

the objective becomes identifying the best possible trade-offs between competing requirements.

Mapping this Pareto frontier accurately requires evaluating many candidate solutions across the design space.

The limiting factor is therefore not whether a design can be simulated.

The limiting factor is:

How many simulations can reasonably be performed?

---

Why Conventional Optimisation Fails

Because high-fidelity simulations remain computationally expensive, traditional optimisation algorithms frequently struggle when applied directly to complex engineering problems.

---

Gradient-Based Methods

Gradient-based algorithms, such as Sequential Quadratic Programming (SQP), navigate the design space by following the local slope of the objective function.

The optimisation process requires evaluating:

$$\nabla f(\mathbf{x})$$

For simulation-based functions, gradients are commonly approximated using finite differences.

For $n$ design variables:

$$\frac{\partial f}{\partial x_i} \approx \frac{f(\mathbf{x}+\Delta x_i)-f(\mathbf{x})}{\Delta x_i}$$

requires approximately:

$$n+1$$

separate simulations.

For high-dimensional engineering problems, this computational scaling quickly becomes prohibitive.

Furthermore, gradient methods assume the design space is:

• Smooth
• Continuous
• Differentiable

Real engineering systems frequently violate these assumptions, the resulting discontinuities and nonlinearities can cause:

• Convergence failure
• Local optimisation around poor solutions
• Sensitivity to initial conditions

---

Evolutionary Methods

Evolutionary algorithms, such as Genetic Algorithms (GA), avoid gradient requirements by exploring populations of candidate solutions.

Evolutionary algorithms explore populations of candidate solutions through selection and stochastic variation:

1. Generate candidate designs.
2. Simulate their performance.
3. Select high-performing solutions.
4. Apply mutation and crossover.
5. Repeat.

Figure 6. Evolutionary optimisation loop.

Evolutionary methods are extremely powerful for:

• Nonlinear systems
• Multimodal problems
• Discontinuous design spaces

However, they require large numbers of evaluations.

A typical optimisation may require:

$$10^3-10^5$$

simulation evaluations.

When each evaluation requires hours of high-fidelity computation, evolutionary optimisation becomes computationally impractical.

---

Design of Experiments

Design of Experiments (DOE) methods attempt to efficiently sample a design space by selecting a predefined set of points.

Common approaches include:

• Grid Search
• Latin Hypercube Sampling
• Sobol Sequences

The objective is to maximise the information gained from a limited number of simulations.

Figure 7. Transition from predefined sampling strategies toward adaptive simulation selection.

However, these methods remain fundamentally non-adaptive.

The sampling strategy is determined before simulations begin.

If simulations reveal that:

• A region contains promising designs
• A region violates physical constraints
• A particular parameter interaction dominates performance

the sampling strategy cannot automatically redirect computational effort.

Large amounts of computational resources may therefore be wasted exploring regions with little engineering value.

---

Towards Intelligent Search

The limitations of conventional optimisation suggest a different perspective.

Rather than treating numerical simulation as an isolated analysis tool, simulation can be considered part of a larger information-generation process.

Different models provide different forms of knowledge:

The challenge is therefore not simply selecting the most accurate available model.

Instead, the problem becomes:

How should computational resources be allocated so that every simulation provides maximum information?

This represents a fundamental change in engineering optimisation.

The objective is no longer:

Find the best design by evaluating everything.

Instead:

Learn the structure of the design space while intelligently deciding which designs deserve expensive evaluation.

---

From Simulation to Information

If simulation is viewed as a source of information rather than merely a source of numerical results, a new class of optimisation strategies becomes possible.

The central question becomes:

Can a computational system learn:

$$f(\mathbf{x})$$

and progressively decide:

• Where high-fidelity simulation is required
• Where lower-order models are sufficient
• Which unexplored regions contain valuable information

This question forms the foundation for the remainder of this series.

---

Chapter Summary

Modern engineering design is fundamentally a search problem.

The engineering objective can be represented as the mapping:

$$\mathbf{y}=f(\mathbf{x})$$

where design variables are transformed through physical models into measurable performance metrics.

Representing this function requires a hierarchy of engineering models:

1. Analytical equations
2. Lumped system models
3. Reduced-order simulations
4. High-fidelity numerical solvers
5. Experimental validation

As demonstrated by the pulse jet case study, increasing model fidelity provides improved physical understanding but introduces severe computational cost.

The fundamental limitation of modern computational engineering is therefore not the ability to simulate complex systems.

It is the inability to simulate every possible design.

---

Limitations of Conventional Optimisation

Traditional optimisation methods struggle because:

The future of computational design therefore lies not simply in:

• Faster computers
• More detailed simulations

but in intelligent systems capable of managing simulation fidelity as a strategic resource.

The solution is not to eliminate high-fidelity simulation, but to use it selectively. The following sections investigate how surrogate models, uncertainty quantification and active learning allow engineering systems to learn the structure of a design space before committing expensive computational resources.