Agenda

What do you do when the simulation that guides your most expensive engineering decisions takes three days to run?

 

In aerospace investment casting, porosity above the acceptance threshold sends a turbine blade worth tens of thousands of euros to scrap. Predicting where that porosity will form means solving heat transfer and fluid flow across the entire casting geometry, which takes hours to days per run on HPC infrastructure. The practical consequence is that design iteration runs at the speed of the solver. Engineers evaluate a handful of process configurations instead of the hundreds a real optimisation would need, and every redesign loop costs days.

At ACCESS e.V., we stopped treating our simulation archive as output and started treating it as training data. PorosAI is the result: an ML surrogate that predicts porosity directly from casting conditions in seconds rather than hours. Optimisation that used to run overnight now runs interactively, cutting design iteration time and shortening product development. It is now in pilot use with foundry engineers at our Techcenter.

In this session, I will walk through the build end to end, with real numbers: how we turned a heterogeneous archive of solidification runs into a supervised training set, and how competing ensemble approaches benchmarked against each other. Purely data-driven features hit a performance ceiling that more feature engineering could not move, until a single physics-guided descriptor, derived from a classical foundry rule from the 1940s, broke it.

 

You will leave with:

No prior knowledge of casting or metallurgy required. The pattern applies wherever expensive simulations create bottlenecks and years of historical runs sit unused.