21 Comments
User's avatar
Damian Hadley's avatar

Dominique, on buildings the hand prediction is usually easier to make than on a machined part, which goes some way to answering Eric. A floor slab spanning between beams behaves closely enough to a simple beam that load times span squared over eight gives the bending moment within the margin you ask for. Under wind, the whole building acts as a cantilever standing up out of the ground, so the overturning at the base is the wind force times the height it acts at. The check I run on every ETABS model is the base reaction. Before the model runs, I multiply each floor's area by its load per square metre and add the floors up on a calculator. A 600 square metre floor carrying about 9 kPa is about 5,400 kN, so a five storey building should come to around 27,000 kN at the base. If the model's total comes out 5,400 kN short, a floor was left unloaded, and if it comes out 5,400 kN over, a load went in twice. I find that before anyone has looked at a stress plot.

Dominique Madier's avatar

Thank you, Damian. This is an excellent illustration. What I like most about your base reaction check is that the size of the discrepancy tells you what went wrong. A shortfall of exactly one floor's load points to an unloaded floor. An excess of one floor's load points to a duplicated load case. The error isn't just detected, it's diagnosed, and all before anyone has been anchored by a stress plot. That is the prediction-first mindset in its most practical form.

Damian Hadley's avatar

Dominique, the base reaction check adds up every load in the model and compares it with the total I worked out by hand, so it catches a load that was left out or entered twice. What it can't catch is a load entered in the wrong place. If the roof plant goes in over the wrong bay, the total at the base is still right, because the same weight is in the model, only sitting over different columns. So I also check a handful of columns on their own. For each one I take its tributary area (half the distance to the next column in each direction), multiply it by the load on each floor above, add the floors up, and compare that with the axial force the model gives at the bottom of the column. When a column is carrying well over that figure, the usual cause is a load drawn over the wrong part of the plan.

Muhammad Ismail's avatar

This is one of the most insightful articles on FEA I have come across. Thank you, Dominique, for articulating so clearly a principle that experienced analysts understand instinctively: software should confirm engineering reasoning, not replace it.

The FEA community would greatly benefit from a practical handbook that connects hand calculations with FEA results across a wide range of examples. Such a reference would be an invaluable learning tool for developing the engineering judgment that ultimately determines the quality of any analysis.

Dominique Madier's avatar

Thank you, Muhammad, for your kind words. You've summarized the article better than I did in its subtitle: software should confirm engineering reasoning, not replace it.

Your handbook suggestion is a strong one, and you've identified exactly what's missing. Plenty of references cover hand calculations, and plenty cover FEA, but very few put the two side by side and explain why they agree, why they differ, and what the difference teaches you. That reconciliation step is where judgment is built, and it's rarely written down.

If you were to use such a reference, which cases would you most want to see covered? Knowing where engineers feel least confident making a prediction would help shape what comes next, whether in the follow-up piece or something more substantial.

Muhammad Ismail's avatar

Thanks, Dominique. My primary interest would be applications in ships, offshore structures, and oil & gas subsea systems, particularly where engineers must reconcile hand calculations with FEA in areas such as local stress concentrations, contact problems, and complex load paths. That said, I'd also be very interested in lessons from automotive and aerospace, where many valuable modelling and validation lessons can often be transferred across sectors.

Nathan Nacer's avatar

Therefore every structural engineer should first master Continuum Mechanics / Solid mechanics with topics like Tensors , deformation (stress and strain ) and linear Elasticity??

Dominique Madier's avatar

Absolutely Nathan. These are the fundamentals a structural engineer should know before to start modeling something.

Antonio Sindoni's avatar

I agree with the sequence you propose Dominique, though I admit that for short time constraints reasons in projects, I generally see engineers creating the FEA model before any hand calcs.

Personally, for same reasons it has happened to me more times to do the same sequence. But I am realising by doing more and more designs that actually time is not a good reason.

It is not a good reason because you will anyway need to spend a lot of time trying to understand what is happening in the model after you got your results.

It is very very difficult to get at first shot a model that gives trustworthy results and you can’t really trust until you have hand calcs to prove it. So why don’t do it beforehand?!

Personally I have been recently trying to change my approach when I create FEA models by creating a very simple models first with the constraints and load I want so to back check by hand calcs the loads applied, reactions and stresses. Once I am sure about those I gradually make little more complex models to continue back checking, until I reach my final model for the project. That is the moment I cannot expect the same results as my hand calcs and I just try to see how load flows through beams and how reactions distribute among supports with respect to how the load is applied.

Dominique Madier's avatar

Absolutely Antonio. You've put your finger on the real economics of it: skipping the hand calculation doesn't save time, it only moves the time to the end of the project, where it costs more. Interpreting a result you have no expectation for is slow, uncertain work, and it usually happens under the most deadline pressure.

The approach you describe, starting with a very simple model and adding complexity step by step, is one of the most effective verification strategies I know. Its strength is that each step changes only one thing. If the results shift unexpectedly, you know exactly where to look, instead of searching through an entire detailed model for the cause.

One small habit makes it even more powerful: before each step, write down what you expect the change to do. Adding a stiffener should pull load toward it. Releasing a rotation should increase the deflection and redistribute the moments. When the model confirms the expectation, you move on with confidence. When it doesn't, you've found either an error or a piece of physics worth understanding.

And I'd argue that your final stage is still prediction, just at a different level. You can no longer expect the hand calculation numbers, but you can predict how load will flow and which supports will pick up the most reaction. That is exactly the kind of judgment the article is about.

Josh Hammer's avatar

I do try to provide supporting hand calcs as a way to build confidence in an FEA, but I usually do them after getting results. Requiring them beforehand is an interesting workflow - one that I’ll likely start using myself. 😊

Dominique Madier's avatar

Thank you, Josh. Doing supporting hand calculations at all already puts you ahead of many analysts, so the change you're considering is a small one in effort but a big one in value.

The reason the order matters is anchoring. Once you've seen the FEA result, it's very hard not to let it shape the hand calculation. An assumption gets adjusted here, an effective width chosen there, and the two numbers end up agreeing because one was steered toward the other. Written down beforehand, the prediction stays independent, and that independence is what makes it a real check.

Eric Cabrol's avatar

I've read a similar argument a few days ago on LinkedIn

(https://www.linkedin.com/feed/update/urn:li:activity:7508065503232446464/) and I'm afraid I mostly disagree not about the intent, which is virtuous, but about the possibility to put it in practice. In my FEA career (about 15 years), I think there would have been very few situations that I could have approached with simple enough approximations for hand calculations to be relevant.

I used to work in the automotive industry, and we don't see cantilever beams that often :)

Dominique Madier's avatar

Thank you for the thoughtful pushback, Eric. I agree the intent means little if it cannot be applied, but I would argue it almost always can, as long as we stop expecting the hand calculation to reproduce the model.

The goal is not to calculate the part. It is to predict the global behavior. Automotive structures rarely look like cantilever beams, but a suspension arm still carries load mostly as a two-force member, a body in torsion still behaves roughly like a closed box section, and a bracket still has to send its load somewhere through a path you can sketch. Even a free-body diagram checking that reactions balance the applied loads, or a mass and 1g check, catches a surprising share of modeling errors.

A prediction can also be qualitative: where the maximum displacement should be, which direction it should go, which member should carry most of the load. If you cannot say that before pressing Solve, you have no way to tell whether the plot is right. In my experience, that is almost never impossible, just rarely asked for.

Adeodatus's avatar

It’s going to be interesting when the inflection point Anthropic worries about comes to pass. Right now human + AI leads to better outcomes than humans alone or AI alone

What happens to accountability and liability when an inflection point occurs, when human in the loop causes more negative outcomes, more errors and is statistically more dangerous/error prone.

This perspective only lasts as long as we are in the Goldilocks zone of human + AI collaboration. And their concerns are warranted, how will society be able to deal with notions like liability and responsibility when the day comes that human in the loop itself becomes a liability and the main source of error. Right now humans do need to error check AI, soon that process of error checking will be where the majority of errors are introduced due to the human element

Dominique Madier's avatar

Thank you, Adeodatus. This is a serious question, and I don't think anyone has a complete answer yet.

The inflection point may well come for certain tasks. But notice that recognizing it requires exactly the skill the article defends. To show that human review now adds more errors than it removes, someone has to judge which results were right and which were wrong. That judgment doesn't disappear when the human leaves the loop. It moves upstream, into validating the system itself.

Aerospace already has a version of this. Autopilots fly more precisely than pilots for most of a flight, yet pilots are still trained to fly by hand, and accountability sits in certification: the system is qualified against evidence, and humans own the definition of its operating envelope. I expect FEA to follow a similar path. Liability will shift from signing off each result to qualifying the tools and defining the problem correctly.

That last part matters most. In my experience, the majority of serious errors are not in the solution but in the problem definition: the wrong loads, the wrong boundary conditions, the wrong assumption about how the structure behaves. An AI can solve a wrongly posed problem flawlessly. Knowing what the answer should roughly be is what tells you the question was wrong.

Adeodatus's avatar

What you said at the end still will be valuable for sure, there’s still value in experience and intuition to know what to expect. I’m not doing anything really exciting, I do some dfm for injection molding and then some mold flow simulations to optimize the design, nothing safety critical or really exciting like FEA can be.

I’m confident AI will be able to do what I do very soon, but before I went for my engineering degree I worked as a plastics processor for a decade. It’s quite often that I over ride the simulation on design decisions just because it’s a) only so good and b) I have experience and intuition to over ride it

That last part is where engineers will still add value, but I have no idea how employers will ever quantify or identify that in candidates

Dr Matt Carrington's avatar

This echos concerns I have been discussing several times this week.

Dominique Madier's avatar

Thank you, Matt. It's reassuring to know these concerns are coming up in other conversations too. It suggests this isn't just one analyst's frustration but a pattern many teams are noticing. I'd be very interested to hear what prompted those discussions. Was it a specific error that slipped through, a hiring or onboarding question, or the growing role of AI in model setup?

Dr Matt Carrington's avatar

Hi Dominique. It's coming from a number of different things at the moment, but largely coming out of stress reviews and the development of Junior Engineers.

Dominique Madier's avatar

Thank you for the info Matt