Guide

When can you trust an AI prediction?

A single predicted number tells you nothing about whether to act on it. An uncertainty-aware model returns a range alongside every prediction, so you can see where it is confident and where it is guessing. That turns a black box into a decision: trust the confident region, and send the rest to the full solver.

What the range reports, how to read it on a real case, and where its meaning stops.

Uncertainty Quantification How STOCHOS works
A DIM-GP prediction with a mean line and one, two, and three sigma confidence bands that stay tight where data exists and widen where the model extrapolates.
Tight where the data is. Wide where it is not. That widening is the whole answer to the question.
Twelve labeled inputs feed a closed model that returns a single bare prediction of 42.7, with no indication of how far to trust the number.
The output looks identical whether the model knows the case or has never seen anything like it.
The starting point

The problem with a point prediction

A model will answer any question you ask it, including questions about designs nothing in its training data resembles, and a bare number gives you no way to tell those two cases apart.

Both answers arrive in the same format, to the same decimal place, with the same apparent authority. One is an interpolation between cases the model has seen many times. The other is an extrapolation into territory it has never been shown.

Nothing in the output separates the two, so the risk does not disappear. It moves to whoever signs the release, who now has to judge the model without any information from it.

What you get instead

What quantified uncertainty actually reports

Alongside the predicted value you get a confidence range that widens where the model has seen little relevant data, which is the signal you use to decide.

Where your data is dense the range stays tight, and the model is telling you it has handled this kind of case before. Where the range opens out it is telling you the opposite, without anyone having to guess which region a candidate fell into.

That makes the output a two-part answer: the estimate, and how much weight it will carry in a review. Uncertainty Quantification covers the two kinds of uncertainty behind that range.

Narrow range

The case sits inside data the model has already fitted. The prediction is worth acting on, and a full solve on it buys little.

Wide range

The model is reaching. Treat the value as a shortlist entry, not a result, and put solver time here rather than everywhere.

A reference case

Reading the range in practice

In an LS-DYNA reference example STOCHOS learned from 32 crash models and predicted a new case with error below 5 percent, after 21 seconds of training on a CPU. The full solver still validates the cases that matter.

32

crash models used as training data

<5%

error on the predicted new case

21s

training time, on a CPU

The speed on its own is not the point. The point is that a prediction arriving in seconds with a stated range lets you sort a candidate list instead of treating every entry the same: act on the confident ones, and spend solver hours on the ones the model itself flags.

These numbers describe that reference case and its setup. What a model does on your data depends on your data.

Inside the trained domain

The range means something. It reports how consistent the model is where it has evidence.

Outside it

The range widens and the honest reading is that the model has left the ground it was trained on. Run the solver.

The limit

Where the boundary sits

Confidence is only meaningful inside the domain the model was trained on, so a design far outside that range is a signal to run the solver, not to trust the number.

It is worth being exact about what confidence claims. It says the model is consistent where it has data. It does not say the answer is proven correct, and no range replaces the validation that confirms a design.

STOCHOS is a predictive model that sits alongside your solver, not in place of it. Used that way, a surrogate model narrows what needs a full run, and the runs you do keep are the ones that decide something.

Common questions

01Does high confidence mean the prediction is correct?

It means the model has seen enough relevant data to be consistent there. Validation against the solver is still what confirms a design.

02What happens outside the training domain?

The range widens, which is the model telling you it is extrapolating. Treat that as a prompt to run the full simulation.

03Does uncertainty slow anything down?

No. With STOCHOS the confidence range is produced together with the prediction, not as a second pass, so there is nothing extra to run.

Related pages
Method

Uncertainty Quantification

Aleatoric and epistemic uncertainty, and why a bare number hides the risk.
Method

Surrogate Modeling

What a surrogate model does, and when it is worth building one.
Product

STOCHOS

The predictive engine that returns confidence with every prediction.

News and Guides collects the rest of the guides and project updates.

Next step

See it on your own data

Bring one dataset. We put a range on the predictions with STOCHOS, on a problem close to yours, and you can see for yourself where it is confident.

Request a Demo How STOCHOS works

Local-first. Your data stays on your infrastructure.

Partners, customers, and research collaborators
AnsysCADFEMSimuTech GroupMEScoTSNEBoschZFGEMUDLRAdler LackeMankiewiczDuluxPlixxentFraunhoferHochschule NiederrheinFUELL Lab AutomationHumotionUniversitaet HamburgRobert Bosch Stiftung