Part IV — The Limits of Autonomy in Computational Design

From Optimisation to Engineering

The previous article concluded with an uncomfortable observation.

An optimisation algorithm does not understand engineering.

It does not recognise physical impossibilities, manufacturing constraints or incorrect modelling assumptions. It simply searches for the maximum of the mathematical model presented to it.

For sufficiently complex optimisation problems this distinction becomes increasingly important. Modern engineering design no longer consists of analysing a single configuration, but of navigating design spaces containing billions or even trillions of feasible parameter combinations. Machine learning, surrogate modelling and active learning provide an effective mechanism for navigating these spaces, but they do not change a fundamental truth.

Every optimisation is only as reliable as the physical knowledge embedded within it.

Consequently, the challenge facing autonomous engineering systems is no longer one of computational capability. Contemporary hardware is capable of performing billions of floating-point operations each second, while modern CFD and finite element solvers routinely resolve phenomena that would have been computationally impossible only a decade ago.

The remaining challenge is one of trust.

How should an autonomous optimisation system determine when a prediction should be believed?

How should it recognise the limits of its own knowledge?

How should computational evidence be balanced against experimental evidence?

These questions extend beyond machine learning. They concern a fundamental principle of engineering: the relationship between models, evidence and physical reality.

Optimisation Exploits Models

Every optimisation algorithm operates according to a remarkably simple principle.

Given an objective function,

f(\mathbf{x}),

the optimiser searches for the design vector,

\mathbf{x}^{*} = \arg\max_{\mathbf{x}} f(\mathbf{x}),

that maximises the predicted engineering performance.

Importantly, the optimiser never distinguishes between physical behaviour and mathematical behaviour.

It assumes that every feature of the objective function corresponds to a real engineering phenomenon.

Unfortunately, engineering models are approximations.

Every governing equation contains assumptions.

Every turbulence model neglects unresolved scales.

Every material model is valid only over a limited operating regime.

Every numerical discretisation introduces error.

Consequently,

f(\mathbf{x}) \neq f_{\text{real}}(\mathbf{x}).

Instead, the optimiser maximises the response predicted by the mathematical model:

\mathbf{x}^{*} = \arg\max_{\mathbf{x}} \left( f_{\text{real}}(\mathbf{x}) + \varepsilon(\mathbf{x}) \right),

where

\varepsilon(\mathbf{x})

represents the cumulative modelling error introduced throughout the simulation pipeline.

From the perspective of numerical simulation this error may appear small.

From the perspective of optimisation it may completely redefine the optimum.

If the modelling error varies across the design space, the optimiser may converge towards regions where the error is largest rather than regions where the true engineering performance is greatest.

The optimiser therefore performs an unexpected role.

Rather than exposing modelling errors, it actively searches for them.

The Pulse Jet Failure

The pulse jet example introduced throughout this series demonstrates precisely this scenario. When an optimiser is allowed to freely explore a pulse jet's exhaust diameter using a simplified flow model, it will confidently converge on an impossibly small exhaust. Because the reduced-order model cannot predict flow choking or boundary layer blockage, it continues to predict higher chamber pressure and increased thrust as the diameter shrinks.

Nothing has malfunctioned. The optimiser behaved correctly. The surrogate behaved correctly. The acquisition function behaved correctly. The failure occurred because the optimisation process successfully exploited a limitation within the mathematical model.

This behaviour is not unique to pulse jets.

Exactly the same phenomenon has appeared throughout computational engineering.

Topology optimisation frequently produces structures that cannot be manufactured.

Aerodynamic optimisation discovers geometries that exploit weaknesses in turbulence models.

Structural optimisation converges towards unrealistically thin members because local buckling has been neglected.

Machine learning has merely made these failures occur more efficiently.

The optimiser is not cheating.

It is revealing precisely where the engineering assumptions begin to fail.

Figure 1. Example of model exploitation in optimisation. The algorithm follows the mathematical behaviour of the model, even when that behaviour no longer represents physical reality.

The Shift in the Engineer's Role

The transition toward Simulation-Driven Design (SDD) and Model-Based Optimisation changes the nature of engineering work.

Historically, much of an engineer's value was associated with their ability to perform complex calculations. The engineer was the person who derived equations, manually evaluated solutions and interpreted the resulting behaviour.

As computational tools advanced, many of these tasks became automated.

Finite element solvers replaced large portions of structural calculation.

Computational Fluid Dynamics replaced many analytical flow approximations.

Circuit simulation software replaced manual analysis of increasingly complex electrical systems.

However, these tools did not remove the need for engineers.

Instead, they changed where engineering expertise was applied.

The engineer became less focused on manually solving individual equations and increasingly responsible for selecting appropriate models, defining assumptions and interpreting results.

Model-based optimisation represents the next stage of this progression.

The engineer is no longer simply operating simulation tools.

They are designing the computational systems that determine which simulations should be performed, how they should be interpreted and when their results should be trusted.

From Simulation Operator to System Architect

A traditional engineering workflow places the engineer directly inside the analysis loop.

A simplified process may appear as:

Define Geometry
      ↓
Create Simulation Model
      ↓
Run Analysis
      ↓
Interpret Results
      ↓
Modify Design
      ↓
Repeat

The engineer personally controls each iteration.

Autonomous optimisation changes this workflow.

Instead of manually evaluating each design variation, the engineer defines the architecture of the exploration process:

Engineering Intent
        ↓
Design Representation
        ↓
Physics Models
        ↓
Optimisation Framework
        ↓
Candidate Designs
        ↓
Validation

The engineer's contribution moves upstream.

Rather than asking:

"What is the pressure distribution for this geometry?"

the engineer increasingly asks:

"What information does the optimisation system require to determine which geometries are worth analysing?"

This represents a significant change in engineering responsibility.

The challenge is no longer performing every calculation.

The challenge is ensuring that the calculations being performed are the correct ones.

Defining the Design Space

One of the most important responsibilities of the engineer is defining the boundaries within which optimisation is allowed to operate.

Every optimisation problem requires a design vector,

\mathbf{x} = \begin{bmatrix} x_1 & x_2 & \dots & x_n \end{bmatrix}^{T},

representing the parameters available for modification.

For the pulse jet example, this may include:

\mathbf{x} = \begin{bmatrix} L_c & D_c & L_t & D_t & \theta & \dot{m}_f & A_i \end{bmatrix}^{T}.

However, the mathematical design space is not automatically equivalent to the physical design space.

A solver may accept a geometry with:

unrealistic wall thickness,
impossible material temperatures,
unstable combustion conditions,
manufacturing limitations,
inaccessible operating regimes.

From the perspective of the optimiser, these designs are simply valid points within the search domain.

The engineer must therefore encode physical reality into the optimisation problem through:

geometric constraints,
material limits,
operating boundaries,
safety factors,
manufacturing restrictions,
valid model regimes.

The quality of an autonomous optimisation system is therefore determined not only by the intelligence of the search algorithm, but by the engineering quality of the problem formulation.

Selecting and Validating the Physics

The engineer's responsibility extends beyond defining geometry.

Every autonomous design system depends upon a chain of physical models.

For the pulse jet, this may include:

Combustion Model
       ↓
Thermodynamic Model
       ↓
Fluid Flow Model
       ↓
Thermal Model
       ↓
Structural Model
       ↓
Performance Prediction

Each component introduces assumptions.

A combustion model may simplify reaction chemistry.

A turbulence model may approximate unresolved flow structures.

A thermal model may assume idealised boundary conditions.

A structural model may neglect manufacturing defects.

The optimisation framework can only reason about the information provided by these models.

It cannot determine whether those assumptions remain valid.

This creates a new engineering responsibility:

model governance.

The engineer must understand:

where a model is accurate,
where it begins to fail,
which assumptions dominate uncertainty,
when additional physics must be introduced.

The most advanced optimisation algorithm cannot compensate for an incorrect physical model.

A perfectly optimised solution to the wrong problem remains the wrong solution.

Knowing When the Surrogate Is Lying

One of the most important skills in future engineering workflows will be recognising when computational models become unreliable.

A surrogate model may provide an apparently excellent prediction,

\hat{y} = f(\mathbf{x}),

but the engineer must also understand the confidence associated with that prediction.

A design located far outside the training distribution may produce a highly confident numerical result despite having little physical support.

This is where uncertainty quantification becomes an engineering requirement rather than simply a machine learning feature.

The engineer must ask:

Has this region of the design space been explored before?
Is the prediction based on interpolation or extrapolation?
Does the result violate known physical trends?
Does the proposed design make engineering sense?
Does an increase in predicted performance correspond to a real physical mechanism?

These questions represent a new form of engineering analysis.

Instead of examining only simulation outputs, engineers must increasingly examine the behaviour of the computational system producing those outputs.

The Physics Guardian

The future engineer is therefore not replaced by autonomous systems.

Their role becomes more fundamental.

As optimisation algorithms become capable of exploring increasingly large design spaces, engineers become responsible for ensuring that those searches remain physically meaningful.

They define the questions.

They establish the constraints.

They select the models.

They validate the assumptions.

They interpret the results.

The computational system may discover a design that no engineer would have manually proposed.

However, only an engineer can determine whether that discovery represents a genuine physical breakthrough or merely an artefact of the assumptions embedded within the model.

The engineer's role shifts from performing every calculation to ensuring that every calculation contributes meaningful knowledge.

The Limits of Autonomy

The development of autonomous engineering systems introduces significant capabilities.

A computational framework capable of combining physics-based simulation, surrogate modelling, uncertainty estimation and active learning can explore design spaces far beyond the practical limits of manual analysis.

However, these capabilities exist within a fundamental boundary.

Autonomous optimisation is exceptionally effective at discovering improved solutions within a defined region of engineering knowledge.

It is far less capable when the problem itself extends beyond that knowledge.

This distinction separates engineering optimisation from engineering invention.

Optimisation Within Known Physics

Most engineering optimisation problems are refinement problems.

The system already exists.

The governing physical principles are understood.

The objective is to improve performance, reduce cost or increase reliability.

Examples include:

improving turbine blade efficiency,
reducing vehicle aerodynamic drag,
optimising heat exchanger geometry,
improving motor electromagnetic performance,
reducing structural mass.

In these cases, the design space contains many possible solutions, but the underlying physics are known.

The role of autonomous optimisation is therefore to efficiently search through existing possibilities.

Mathematically, the problem is well defined.

A governing model exists,

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

and the optimisation process attempts to identify

\mathbf{x}^{*} = \arg\max_{\mathbf{x}} \mathbf{f}(\mathbf{x}).

Although computationally difficult, the problem is fundamentally one of exploration. The challenge is entirely computational.

The Problem of Novelty

The situation changes when engineers enter unexplored physical regimes.

Consider the development of a completely new propulsion concept.

The difficulty is not simply that the number of possible geometries is large.

The difficulty is that the relationship between design variables and performance is not yet known.

The function itself is uncertain.

Instead of

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

the problem becomes

\mathbf{y} = ?

The engineering challenge is no longer searching through a known landscape.

It is discovering what landscape exists.

A surrogate model cannot learn behaviour that has never been observed.

A reinforcement learning agent cannot optimise against a physical mechanism that is absent from its environment.

An active learning system can determine where additional information is valuable, but it cannot independently generate entirely new physical understanding.

The Difference Between Discovery and Optimisation

This distinction is important because the terms are often used interchangeably.

Optimisation asks:

"Given our current understanding of the system, what is the best design?"

Discovery asks:

"What new understanding is required to create a new class of design?"

These are fundamentally different engineering activities.

Autonomous optimisation excels at the first.

The second still requires human creativity, intuition and scientific investigation.

Historically significant engineering developments often emerged from conceptual breakthroughs rather than exhaustive search.

Examples include:

the transition from propeller aircraft to jet propulsion,
the development of semiconductor technology,
the invention of composite materials,
the emergence of additive manufacturing.

These advances were not simply the result of searching a predefined design space.

They required engineers to redefine the design space itself.

The Problem of Sparse Knowledge

Even within established engineering domains, autonomous systems may encounter regions where available data is insufficient.

A surrogate model trained on existing designs naturally inherits the limitations of those designs.

For example, an aerodynamic optimisation system trained primarily on conventional aircraft geometries may confidently optimise within that domain.

However, if presented with a radically different configuration, such as a novel blended-wing concept, the model may have limited ability to determine whether the design represents:

a genuine improvement,
an unexplored opportunity,
or an unrealistic prediction.

The challenge is not simply uncertainty.

It is uncertainty about the structure of the problem itself.

Traditional uncertainty quantification asks:

"How uncertain is the prediction?"

A more fundamental question is:

"Are we using the correct model of reality?"

This is a question that cannot always be answered statistically.

The Human Role at the Boundary of Knowledge

The limitations of autonomous optimisation therefore define rather than diminish the role of the engineer.

When operating inside known physics, autonomous systems can dramatically accelerate exploration.

When approaching unknown physics, human engineers remain responsible for creating the new knowledge upon which future optimisation depends.

The engineer provides:

conceptual innovation,
physical intuition,
identification of missing phenomena,
selection of new modelling approaches,
interpretation of unexpected results.

In this sense, autonomous engineering is not a replacement for creativity.

It is an amplifier of creativity.

By removing the computational burden of evaluating thousands of conventional alternatives, engineers can dedicate more effort towards asking deeper questions:

Are we solving the correct problem?
Are the assumed constraints necessary?
Is there a better physical mechanism available?
Are we exploring the right design space?

The greatest engineering advances rarely come from searching faster.

They come from asking better questions.

The Future Boundary

The future of engineering design will therefore not consist of fully autonomous systems independently creating technology.

Figure 2. The future of computational engineering is a collaborative loop where human judgement defines direction while computational systems expand exploration capability.

Instead, it will involve increasingly sophisticated partnerships between human reasoning and computational exploration.

Machines are exceptionally effective at:

evaluating large numbers of possibilities,
identifying complex relationships,
detecting patterns within data,
managing computational resources.

Humans remain essential for:

defining meaningful objectives,
recognising unexpected opportunities,
questioning assumptions,
creating new conceptual frameworks.

The most powerful engineering systems will therefore not remove humans from the design loop.

They will expand the scale at which human engineering judgement can operate.

The Amplification of Engineering Judgement

Engineering simulation was created to reduce uncertainty in physical design.

By converting physical behaviour into mathematical models, engineers gained the ability to evaluate concepts before committing to expensive manufacturing and experimental testing. Analytical methods, numerical simulation and multiphysics analysis progressively expanded the complexity of systems that could be investigated computationally.

However, this advancement created a new engineering challenge.

As simulation capability increased, so did the size and complexity of the design spaces being explored.

Modern engineering systems may contain:

hundreds of coupled design variables,
multiple competing objectives,
nonlinear physical interactions,
multidisciplinary constraints.

The difficulty is no longer simply determining whether a design can be simulated.

The difficulty is determining which designs deserve simulation.

This has been the central theme throughout this investigation.

The Evolution of Computational Engineering

Throughout this series, the central challenge has remained the same: how can engineers extract reliable knowledge from increasingly complex physical systems while operating within practical computational limits?

Part I established the foundation of simulation-driven engineering design by examining the relationship between physical models, computational cost and engineering accuracy.

The pulse jet case study demonstrated that no single modelling approach is universally appropriate. Analytical models provide rapid insight into dominant physical mechanisms. Reduced-order models enable efficient exploration across large design spaces. High-fidelity multiphysics simulations provide detailed predictions required for validation and final engineering decisions.

The objective is therefore not to maximise simulation fidelity at every stage of development.

It is to select the appropriate level of physical representation required to obtain the greatest engineering insight from available computational resources.

Part II addressed the computational limitation created by expensive simulation.

As engineering systems become increasingly complex, the number of possible design configurations rapidly exceeds the practical limits of direct simulation. Surrogate models provide a solution by constructing approximate representations of expensive physics, allowing large regions of the design space to be explored without repeatedly performing high-fidelity analysis.

Neural networks provide the capability to represent highly nonlinear engineering relationships, while Gaussian Process models introduce uncertainty estimation and quantify confidence in those predictions.

Together, these approaches transform simulation data into a continuously improving representation of engineering knowledge.

However, an accurate prediction alone is insufficient.

An intelligent engineering system must also determine where additional information is required and which simulations provide the greatest value.

Part III extended this concept by introducing intelligent simulation selection and computational resource allocation.

Active learning transforms optimisation from a process of simply searching for the best design into a sequential decision-making problem: determining which simulation should be performed next to maximise engineering knowledge.

Acquisition functions balance predicted performance, uncertainty and information gain, allowing computational resources to be directed towards regions where additional analysis is most valuable.

Multi-fidelity optimisation extends this principle by recognising that engineering simulations exist within a hierarchy of cost and accuracy. The optimisation framework must therefore determine not only which design should be investigated, but also which level of physical modelling is appropriate.

However, this progression introduces a fundamental limitation.

An optimisation algorithm can efficiently navigate a design space.

It cannot determine whether that design space accurately represents reality.

This represents the boundary between computational optimisation and engineering judgement.

The future of autonomous engineering is therefore not defined by increasingly powerful search algorithms alone.

It depends upon creating systems capable of recognising uncertainty, validating assumptions and understanding the limits of their own predictions.

Computational methods can determine where to search.

Physics determines what is possible.

Engineering judgement determines whether the answer can be trusted.

The ultimate objective of autonomous engineering is therefore not a system that replaces human decision-making.

It is a system that amplifies it.

By allowing computational tools to perform the repetitive process of exploring vast numbers of possibilities, engineers can focus on higher-level activities:

defining meaningful problems, developing new concepts, identifying missing physics, interpreting unexpected behaviour, making informed engineering decisions.

The future of engineering design will not be defined by choosing between human intelligence and artificial intelligence.

It will be defined by integrating them effectively.

Physics provides the foundation.

Simulation provides the evidence.

Machine learning provides the ability to navigate complexity.

Engineering judgement provides the direction.

Together, these components form a design methodology capable of exploring systems that would otherwise remain inaccessible due to computational limitations.

The purpose of advanced Model-Based Optimisation is not to automate the engineer.

It is to expand what engineers are capable of designing.