Utilization prediction by a machine-learned model

This functionality is provided for selected parametric connection templates included in the IDEA StatiCa predefined set. In version 26.0, only a limited number of templates support the prediction, but the coverage will be gradually expanded in future versions.

For supported templates, the predicted utilization is displayed directly in the scene, providing immediate feedback on the expected capacity of the current configuration. The displayed value represents an estimate generated by a machine learning model.

The prediction method

The predicted utilization is based on a machine learning model trained on a large dataset of pre-calculated connection variants. For each supported parametric template, tens of thousands of models with different parameter combinations were automatically generated.

All these models were fully calculated using standard IDEA StatiCa Connection analysis. Based on the results, a machine learning model was trained to predict the template's utilization for new parameter combinations that were not explicitly calculated beforehand.

The goal of this approach was to provide a fast estimate of the connection utilization without the need to run a full calculation after every parameter change.

How the prediction is evaluated

For each supported parametric template, the machine learning model considers the template's geometry and parameters, along with the load components applied during training.

When the user modifies a template parameter (e.g., plate thickness, bolt size, or edge distances), the predicted utilization is updated immediately. This allows the user to quickly assess whether the current configuration is closer to or further from the expected capacity.

The benefits in IDEA StatiCa Connection

Thanks to the predicted utilization, the user can optimize the connection configuration without repeatedly running the full calculation. The prediction provides instant feedback during parameter changes.

Once the user is satisfied with the predicted utilization, the final verification must always be performed by running the standard stress/strain calculation. Only the calculated result represents the actual design check according to the selected code.

This approach significantly accelerates the design process, as it is no longer necessary to verify every intermediate design variant by calculation.

Displayed prediction in the scene

If multiple load effects are defined in the project, the predicted utilization displayed in the scene always corresponds to the load effect in which the connection reaches the highest utilization. Predictions for other load effects are not shown simultaneously.

Similarly, when two or more parametric templates are used within a single connection model, the predicted utilization displayed in the scene represents the most utilized template, not the utilization of the joint as a whole, because their interaction is not taken into account.

FIG

Predicted utilizations for individual templates can be viewed by selecting each template separately. This allows the user to assess the estimated utilization of each template independently, while keeping in mind that the displayed value does not represent the combined behavior of the entire connection model.

FIG

Limitations

The predicted utilization is intended to provide a fast estimate during connection design. When interpreting the displayed value, it is important to understand the assumptions used during training of the machine‑learning model and the inherent limitations resulting from them.

The following limitations apply to the predicted utilization:

 1  Load considerations

• Limited load components — The training of each template included only two load components (e.g., Vz and My), with anchoring cases trained for three components. Other load components are not considered for prediction - if these components are decisive for the connection behavior, the predicted utilization may differ significantly from the calculated result.
• Connected member only — The prediction considers loads applied to the connected member only. Loads acting on the supporting member are not included. Where the supporting member is significantly loaded, the calculated result may be substantially less favorable than the prediction.

2  Modeling assumptions

• Butt (CJP) welds assumed — The models were trained assuming butt (CJP) welds. Weld size is therefore not reflected in the predicted utilization. If weld capacity is decisive, the prediction may differ significantly from the calculated result. Weld autodesign with overstrength, for full-strength, minimum ductility, or capacity estimation is available in the project item context menu.
• Default project settings required — The predicted utilization offers reliable results only if the project settings match those used during training. This means that default Project settings must be used, in particular, safety factors, mesh settings, default lengths, and segment division.

3  Availability

• Only for EN and AISC design codes.
• Only when Stress–Strain analysis is selected.
• An internet connection is required.
  

4  Template modifications

• Developer tab & structural scheme changes — If the user modifies the template in the Developer tab (e.g., removing parameters, changing parameter limits) or selects a different supporting member, the prediction becomes unavailable. In such cases, the algorithm cannot control how much the modified connection differs from the original trained template.
• Geometric changes — Some geometric changes are not yet reflected in the predicted utilization. For example, changing the angle or axis eccentricity of the connected member is not taken into account, and in such cases, the predicted utilization may differ significantly from the calculated result.
• Additional operations — If an additional operation is added to the template (e.g., a stiffener), the predicted utilization will still be displayed, but the effect of that operation is not included in the prediction.
  

5  Multiple templates

• When multiple templates are used within one connection model, the utilization of each template can be predicted separately. However, their mutual interaction is not considered.

Final remark

The predicted utilization should always be understood as an estimate. Machine learning tools cannot replace a full structural calculation performed by an engineer.

The prediction can provide significant time savings during connection optimization by allowing fast comparison of different parameter configurations.

The listed limitations will be gradually reduced as the machine learning process and training datasets are further improved. At the same time, the number of supported parametric templates with prediction functionality will continue to increase.