Development of an AI tool for predicting the performance of steel columns subjected to a blast load

Finite Element Analysis (FEA) of building structural columns subject to a blast loading are time consuming and even a simple single column model can take several hours to build, analyze and post process. These types of high-fidelity analyses are typically done during the detailed design phase of a new building when there is sufficient time for them to be carried out. However, by the detailed design phase the design is well established, and it is harder to accommodate design changes. A quick assessment tool that can be used during the concept design phase would be invaluable in allowing architects and engineers to work out the required column size while the design is still flexible.

To this end we are developing an AI tool that can be used for assessing the blast performance of steel columns. The tool uses a neural network surrogate model of the full fidelity FEA analysis, and to date, predicts analysis results for steel columns with I-section profiles. Outputs that are predicted by the surrogate include likelihood of failure, residual column capacity after a blast event, and a basic characterization of damage level from the blast, alongisde confidence levels in each of these predictions.

We used the LS-DYNA finite element analysis software to run a suite of column blast analyses for a range of column sizes, charge masses, standoffs, and column orientations.

The raw results from these models were post-processed to categorize the performance of each column (failure, damage, remaining load capacity). These were then fed into a machine learning algorithm to develop the AI tool. Our surrogate can both predict the performance as well as the accompanying uncertainty in its prediction, thereby overcoming the black-box nature of AI predictions.

The tool is accessed through an API allowing it to be used from a simple web interface or linked to from other software such as Excel or Rhino 3D