Skip to content
Queensferry

Computational Engineering

Integrated projects

Briefs rather than exercises. Each one has a client question, decisions with no single right answer, and an acceptance list describing what a defensible submission contains.

There are 15 projects, roughly 65 hours of work between them, and between them they draw on all 21 modules of the course. None draws on fewer than four — a project that needs only one module is an exercise wearing a brief’s clothes.

None of them has an answer key, and that is deliberate. A practice question has an answer the course can check; a project has a method the course can check. So each acceptance list names the evidence a reviewer would look for, and every one of them requires the prediction to have been recorded before the model was run — because that is the one step which cannot be done afterwards.

Each project also names the trap it exists to spring. That is unusual and it is not generosity: the traps are all cases where the wrong answer looks entirely reasonable, and a trap you have been warned about is still one you have to actively avoid.

The sequence

Every project is worked with the sixteen-step sequence from Module 21. Eight of the sixteen steps happen before any model exists, and the two most often skipped — recording a prediction and writing a verification plan — are both free.

Try it

Integrated project workspace

Choose a brief and work the sixteen steps. The prediction cannot be entered once the analysis has run, because a prediction made afterwards is not one.

Project brief

Steps completed
0 of 16
Before any model (1–8)
0 of 8
Prediction recorded
no
Verification plan recorded
no

Once this is pressed, step 8 locks. That is the whole mechanism: a prediction written after seeing the result is a search for reasons the number might be right.

Six-storey office, 12 m grid proposed

8 000 m² over six storeys. The architect proposes a 12 m column-free grid for letting flexibility. The client has committed publicly to an embodied-carbon target. Floor zone agreed at 1 100 mm including services.

The engineering question, in one sentence.

Several genuinely different forms, not variations of one.

What you will vary, and what you are fixing.

Where geometry comes from, where results go, what is lost.

Whole or component, 1D/2D/3D — and what it will NOT answer.

From the required outputs.

List the nonlinear features first.

Estimate the answer NOW, before the model exists.

With the diagnostics run as you go.

Written before results arrive. What check, against what.

Which inputs it depends on, with ranges that have a basis.

Where the question is 'which of these'.

Same method for every option; test the ranking.

Any legitimate role? Does a deterministic method exist?

Purpose, assumptions, exclusions, limitations, sign-off.

What you advise, on what evidence, what would change it.

8 of the first eight steps are still empty. Those cost about a morning between them and decide almost everything; steps 9 onwards cost weeks and can only implement what they decided.

What none of these steps supplies

  • An expectation of what the structure does. Every check depends on having one, and it comes from Structural Behaviour rather than from any procedure.
  • The judgement that the brief has already given away the answer — which is what the module's worked project turned on.
  • How much checking is enough. Module 13 said it should follow the consequence; judging the consequence is not on any list.

What this shows: Ten of the sixteen steps come before the first result is read — and the two that cost nothing, prediction and the verification plan, are the two most often skipped.

One member

The transfer beam that is too stocky for a line model

about 3 hours

A residential block has a column that must stop at second floor. The transfer beam is 900 mm deep over a 3.6 m clear span, carrying four storeys of column load. The architect wants the depth reduced.

How much does the transfer beam deflect, and does the answer depend on something the usual model discards?

Decisions you must make and justify

  • Whether a line model is adequate at this span/depth ratio
  • Whether to include shear deformation, and how
  • What the supports actually are, given the beam sits on two columns of finite width
  • Whether reducing the depth changes which model is appropriate

A defensible submission contains

  • The span/depth ratio is computed and the shear share of the deflection is quantified rather than asserted
  • A model-purpose statement names what would make the model wrong
  • The deflection is predicted by hand before any model is run, and the prediction is recorded
  • The support width is either modelled as an offset or excluded with a stated reason
  • The answer states which model was used and what it discarded

The trap

At L/d around 4 the line model omits several per cent of the deflection, and reducing the depth makes it worse rather than better — the direction most people expect is backwards.

Draws on: 6. From physical structure to analytical model, 7. Finite-element building blocks, 9. How the stiffness method works, 12. Diagnosing computational models, 13. Validation, verification and model assurance

A wall the model says can carry tension

about 3 hours

A masonry gable wall is modelled with shell elements for a wind check. The results show tension across bed joints near the corners.

What does the model's tension mean, and what should be designed for?

Decisions you must make and justify

  • Whether the model's linear elastic assumption is admissible here
  • How to establish what the wall can actually carry
  • Whether the load path in the model exists in the wall
  • What to do about the corners specifically

A defensible submission contains

  • The middle-third check is applied and the region where the resultant leaves it is identified
  • The model's tension is explicitly identified as a load path the material cannot provide
  • An alternative load path is proposed and checked, or the wall is restrained so one exists
  • The limitation of the linear model is recorded rather than worked around

The trap

No numerical check finds this. The arithmetic is correct throughout; what is wrong is that the material cannot do what the analysis assumed, and only looking at the sign of the stress reveals it.

Draws on: 6. From physical structure to analytical model, 8. Modelling structural systems, 11. Nonlinear and explicit analysis, 12. Diagnosing computational models, 13. Validation, verification and model assurance

One structure

A portal frame whose connections are neither rigid nor pinned

about 4 hours

A 14 m single-bay portal has been designed assuming rigid eaves connections. The fabricator proposes a simpler bolted detail whose rotational stiffness is around twice the rafter's EI/L.

Does the frame still work with the proposed connection, and which member changes most?

Decisions you must make and justify

  • How to represent a semi-continuous connection in the model
  • Which design checks the change affects, and in which direction
  • Whether to accept the detail, require a stiffer one, or resize the rafter
  • What stiffness to design the connection for, given it is now a governing parameter

A defensible submission contains

  • The eaves and midspan moments are computed at the proposed stiffness, not interpolated
  • The free bending moment check — eaves plus midspan equals wL²/8 — is shown for at least two stiffnesses
  • The rafter and the column are both checked, and the direction of change for each is stated
  • The connection stiffness appears in the model record as a governing assumption with a source

The trap

Assuming rigidity is conservative at the eaves and unconservative at midspan. The two errors land on different design checks and do not cancel.

Draws on: 7. Finite-element building blocks, 8. Modelling structural systems, 9. How the stiffness method works, 10. Second-order, buckling and dynamic methods, 13. Validation, verification and model assurance

Is this floor a vibration problem?

about 4 hours

A 9 m span composite office floor is being value-engineered from 450 mm to 400 mm deep beams. The client asks whether it will still be comfortable.

Does the shallower floor fall into the range where walking excitation matters, and how much does that conclusion depend on assumptions nobody has verified?

Decisions you must make and justify

  • What mass to use, and which load case it comes from
  • Whether to run a dynamic model at all, and what would justify the cost
  • What damping to assume and whether the answer survives its plausible range
  • Whether composite stiffness at vibration strains is the same as at ultimate

A defensible submission contains

  • The 18/√δ estimate is computed first and used to decide whether a dynamic model is warranted
  • The mass is unfactored and its composition is stated
  • Damping is swept across its plausible range, and the verdict at the low end is reported
  • If the verdict changes within the damping range, that is said plainly rather than averaged away

The trap

Reusing the ultimate-limit-state mass because that load case already exists. Too much mass gives too low a frequency, which is conservative and can drive a redesign that was never needed.

Draws on: 6. From physical structure to analytical model, 8. Modelling structural systems, 10. Second-order, buckling and dynamic methods, 14. Sensitivity and design-space exploration, 21. The computational engineer in practice

The truss chord that fails on a stress it was not checked for

about 4 hours

A 30 m roof truss has been designed as pin-jointed. The chords are continuous through the panel points, welded, and a checker has queried whether secondary bending matters.

How much bending do the continuous chords pick up, and does the combined stress change any member's verdict?

Decisions you must make and justify

  • How to model continuity while keeping the web members pin-ended
  • Whether the secondary moment is significant relative to the axial stress
  • Whether upsizing the chord helps or hurts
  • How to detail the joints if it does matter

A defensible submission contains

  • The truss is modelled both ways and the axial forces are shown to agree closely
  • The secondary moment is reported as a share of the chord's peak stress, not in isolation
  • The effect of increasing the chord stiffness is computed rather than assumed
  • A combined axial-plus-bending check is made for the governing chord

The trap

Upsizing the chord to fix a secondary bending problem makes it worse. Bending follows stiffness, and the joint rotations being imposed do not change.

Draws on: 7. Finite-element building blocks, 8. Modelling structural systems, 9. How the stiffness method works, 12. Diagnosing computational models, 13. Validation, verification and model assurance

The peak moment that grows with every mesh

about 4 hours

A flat slab has been modelled with shell elements on point-supported columns. The peak hogging moment over a column has doubled at each mesh refinement and the design team wants to know what value to use.

What moment should the slab be reinforced for, and why is the peak not it?

Decisions you must make and justify

  • How to tell a singularity from a stress concentration
  • Whether to model the column as a patch or keep the point support
  • How to obtain a mesh-independent design moment
  • How wide a design strip to integrate over

A defensible submission contains

  • The refinement sequence is tabulated and its behaviour is classified as converging or singular, with the reasoning shown
  • The total static moment across the panel is computed by hand and compared with the integrated model moments
  • The design moment is mesh-independent and this is demonstrated at two mesh sizes
  • The choice between patch support and point support is stated as a decision

The trap

Refining until the moment converges. It will not — a point support in plate theory has an unbounded moment, and each refinement is a more accurate solution of a model containing an infinity.

Draws on: 6. From physical structure to analytical model, 8. Modelling structural systems, 12. Diagnosing computational models, 13. Validation, verification and model assurance

A canopy where the shape is an output

about 5 hours

A cable-supported entrance canopy is to span 18 m. The architect has drawn a shape; the engineer suspects the shape should be found rather than drawn.

What shape does this canopy want to be, and how far is the drawn shape from it?

Decisions you must make and justify

  • Whether to form-find or to analyse the drawn shape
  • What to hold fixed and what to let the form finding decide
  • How to present a found shape to an architect who drew a different one
  • What the drawn shape costs in material, if it is kept
  • Whether to automate the iteration, given how many shapes will be tried

A defensible submission contains

  • A parametric model is built in which the shape is generated by rules rather than drawn
  • The decision to automate the iteration — or not — is justified by counting development, verification and maintenance against the runs saved
  • The found shape is compared with the drawn shape on a stated objective
  • The cost of keeping the drawn shape is quantified rather than described
  • The recommendation acknowledges that the architect's shape may still be the right answer for reasons outside the model

The trap

Treating a form-finding result as the answer. It is the answer to the objective that was stated, and the drawn shape may be serving an objective nobody wrote down.

Draws on: 3. Parametric thinking, 4. Programming and automation for engineers, 15. Structural optimisation, 16. Optimisation algorithms, 21. The computational engineer in practice

How fixed is the base?

about 4 hours

A four-storey braced frame's columns are on pad foundations. The model assumes full fixity at the bases. The geotechnical report gives a modulus of subgrade reaction with a wide range.

Does the design change between fixed and pinned bases, and if so, what soil stiffness is needed to justify the assumption?

Decisions you must make and justify

  • How to represent base compliance
  • What spring stiffness the soil actually gives
  • Whether the difference between fixed and pinned changes any member size
  • What to do if it does and the soil stiffness is uncertain

A defensible submission contains

  • The frame is analysed with fixed and with pinned bases, and the results compared on the quantities that govern
  • The spring stiffness is derived from the geotechnical parameter, with the derivation shown
  • A sensitivity study across the report's stated range is presented
  • If the design changes within the range, soil stiffness is recorded as a governing assumption

The trap

A default spring stiffness is effectively rigid. Base rotation of zero is the signature, and it shortens the period and attracts moment to the base — neither of which is a safe accident.

Draws on: 7. Finite-element building blocks, 8. Modelling structural systems, 10. Second-order, buckling and dynamic methods, 14. Sensitivity and design-space exploration, 13. Validation, verification and model assurance

An optimised truss that has no reserve anywhere

about 4 hours

A size optimisation of a 24 m truss has returned a design in which every member is at 100 % utilisation. The design team is pleased.

Is this a good design, and what has the optimisation not been told?

Decisions you must make and justify

  • Whether uniform full utilisation is a virtue or a warning
  • What constraints were omitted from the optimisation
  • How the design behaves under a load case the optimiser never saw
  • Whether to accept, constrain and re-run, or discard the result

A defensible submission contains

  • The optimised design is checked under at least one load case the optimisation did not include
  • Constraints the optimiser was not given — buildability, member size steps, robustness — are listed
  • The consequence of a single member being under-strength is examined
  • The recommendation says what the optimisation optimised, in those words

The trap

Exactly critical everywhere means no reserve anywhere. An optimiser removes every margin it was not told to keep, and a design with no redundancy is the natural output of an objective that never mentioned redundancy.

Draws on: 14. Sensitivity and design-space exploration, 15. Structural optimisation, 16. Optimisation algorithms, 13. Validation, verification and model assurance, 12. Diagnosing computational models

A whole building

What the grid costs in carbon

about 5 hours

A six-storey office of 8 000 m² is at concept stage. The architect has proposed a 12 m clear-span grid; the structural engineer thinks 7.5 m would be substantially better.

How much embodied carbon does the grid decision move, compared with the material decision, and is the ranking robust?

Decisions you must make and justify

  • Which schemes to compare, and on what objectives
  • Whether to compare on total carbon, intensity, mass, or a combination
  • How to present a result whose factors have not been verified
  • What to recommend when the lowest-carbon option is not the shallowest

A defensible submission contains

  • The spread from scheme choice and the spread from span are both computed and compared
  • The ranking is tested against a stated perturbation of every factor, and the number of flips is reported
  • Every carbon figure is presented with its provenance, and none is presented as verified
  • The recommendation names what would change it

The trap

Comparing materials while holding the grid fixed. The grid moves the number more than the material does, and a material comparison at the wrong grid answers a smaller question than the one that was asked.

Draws on: 14. Sensitivity and design-space exploration, 17. Multi-objective design and Pareto decisions, 20. Computation for low-carbon structural design, 21. The computational engineer in practice

How much lateral load actually reaches the core

about 5 hours

A 12-storey building has a concrete core and a perimeter moment frame. The core has been modelled as a single stick with gross section properties, and it is attracting 85 % of the wind load.

Is the core-to-frame load split credible, and what would change it?

Decisions you must make and justify

  • Whether the stick model over-states the core's stiffness
  • Whether shear deformation and the openings need modelling
  • How the floor diaphragms distribute load, and whether the rigid assumption holds
  • Which of core or frame is unconservative if the split is wrong

A defensible submission contains

  • The core's bending and shear deflection components are both computed and compared
  • The effect of admitting the openings is quantified
  • The diaphragm assumption is stated and tested against the plan proportions
  • The load split is reported as a range, and the frame is checked at the unfavourable end of it

The trap

A bending-only stick model can over-state a squat core's stiffness by a factor of two. Load then goes to the core in the model and to the frame in the building, and only one of those is conservative.

Draws on: 6. From physical structure to analytical model, 7. Finite-element building blocks, 8. Modelling structural systems, 12. Diagnosing computational models, 14. Sensitivity and design-space exploration

A scheme carried through, end to end

about 12 hours

A six-storey office building, 8 000 m², is at RIBA Stage 2. You have the architectural grid, the floor loading and a requirement to report on structural options.

Which structural scheme should be recommended, and on what evidence?

Decisions you must make and justify

  • Which schemes to carry forward and which to eliminate, and why
  • What model to build for each, and what each model may then be asked
  • What to optimise and what to leave alone
  • How to present a recommendation between options that are not commensurable

A defensible submission contains

  • All sixteen steps of the Module 21 sequence are worked, in order
  • The prediction is recorded before any model is run, and it is compared with the result afterwards
  • A verification plan is written before the analysis and its checks are carried out
  • Options are presented as a trade-off with a stated decision and a stated reason, not as a single score
  • Every carbon figure carries its provenance and none is presented as verified
  • The submission says what would change the recommendation

The trap

Building the model first. Eight of the sixteen steps happen before any model exists, and the two most often skipped — the prediction and the verification plan — are the two that are free and cannot be done afterwards.

Draws on: 1. What computational engineering is, 2. Computational design and creativity, 3. Parametric thinking, 5. Digital engineering workflows, 6. From physical structure to analytical model, 7. Finite-element building blocks, 8. Modelling structural systems, 9. How the stiffness method works, 13. Validation, verification and model assurance, 14. Sensitivity and design-space exploration, 17. Multi-objective design and Pareto decisions, 20. Computation for low-carbon structural design, 21. The computational engineer in practice

Working practice

Auditing a model you did not build

about 3 hours

A model arrives from a sub-consultant with a set of results and a request to countersign. There is no model record and the exchange route is a coordinate table.

What can be established about this model, and what must be re-created before anything is signed?

Decisions you must make and justify

  • What to check first, given limited time
  • Which of the twelve model attributes the exchange route could not have carried
  • Whether to check the model or rebuild it
  • What to say in writing about what has and has not been established

A defensible submission contains

  • The total applied load is summed by hand and compared with the reactions
  • The attributes the route cannot carry are listed, and each is either verified independently or declared unverified
  • At least one member is checked by an independent hand calculation
  • The written response distinguishes what was checked from what was assumed

The trap

A clean solve with no warnings is not evidence of a clean transfer. Three of the six common exchange losses leave the model solving and wrong, and one makes the structure look safer.

Draws on: 5. Digital engineering workflows, 12. Diagnosing computational models, 13. Validation, verification and model assurance, 7. Finite-element building blocks

An AI tool recommends a section

about 2 hours

A design assistant suggests a shallower beam than the engineer chose, at 94 % confidence, citing comparable spans in its training data. A deterministic check is available and has not been run.

May this recommendation be acted on, and what would have to be true for it to be?

Decisions you must make and justify

  • What the confidence figure actually describes
  • Whether the recommendation is inside or outside the range the model was trained on
  • Who is accountable if it is wrong
  • What record the decision needs, whichever way it goes

A defensible submission contains

  • The confidence figure is correctly described as a fit to training data rather than a probability of being right
  • The deterministic check is run and its result compared with the recommendation
  • The decision is recorded with a named reviewer and a stated reason
  • The submission states what would make the recommendation wrong

The trap

High confidence with plausible evidence and a senior reviewer still does not clear the bar. Where a deterministic check exists, it must have been run — confidence is not a substitute for a calculation that could have been done.

Draws on: 18. AI and machine learning for structural engineers, 13. Validation, verification and model assurance, 19. Digital fabrication, cloud and connected workflows, 21. The computational engineer in practice

Designing a workflow for a real project

about 3 hours

A project is starting with four consultants, a contractor and a fabricator. You are asked to set out how structural information will move between them.

What are the handovers, who owns each, and what will be lost at each one?

Decisions you must make and justify

  • Which exchanges are needed and which are habit
  • What route each handover uses and what that route cannot carry
  • Who owns each handover, as distinct from who performs it
  • What the handover note must say at each step

A defensible submission contains

  • The workflow is drawn as handovers rather than as software
  • For each handover, what is added, what is lost and who owns it is stated
  • The attributes that survive no handover are identified and a plan for each is given
  • A specimen handover note is written for the most consequential exchange
  • The position on sending project data to external services is stated

The trap

Describing the workflow as a list of packages. The software is not where things go wrong; the arrows between them are, and they are the part nobody owns.

Draws on: 5. Digital engineering workflows, 6. From physical structure to analytical model, 13. Validation, verification and model assurance, 19. Digital fabrication, cloud and connected workflows

Before you submit

The model assurance workflow is the checklist. If a project’s acceptance list and that page disagree about what a submission needs, the acceptance list is the more specific and it wins.

This course teaches analysis and modelling and gives no design resistance. Any project that reaches a member size has reached it by a route the course has not verified, and the design must be made to the current codes by someone who has.