Building Confidence in the Early Stages of Foundry Design
Teaser
A Fast Solidification Assessment Ahead of Full Process Simulation
Bringing a New AI Perspective to Foundry Engineering
Eugen Miknevič has moved from foundry specialist to AI engineer. Following almost two decades in manufacturing and metallurgy, he is now pursuing a new challenge: developing the software he had always wished to have during his own time on the shop floor.
His background brings together extensive industrial knowledge – spanning all major casting processes, CAE simulation and physical metallurgy – with contemporary AI engineering. This combination enables him to connect heavy industry with advanced artificial intelligence. While others may see complex and untidy shop-floor data, he recognises structured data pipelines that are ready for LLMs and vector databases.
He is currently strengthening and broadening his AI expertise through Turing College's intensive, full-time AI Engineering programme. During his studies, he met with Foundry-Planet to present his new software, FoundryFlash.
FoundryFlash: Greater Confidence at the Start of the Design Process
During quotation or at the beginning of the design process, foundry engineers often need a quick answer to a practical question: will the casting solidify in the direction of the risers?
Full process simulation is still required for final approval. However, an initial comparison of different feeding concepts does not always call for the same level of detail.
This distinction is the starting point for FoundryFlash, which Miknevič is developing as a browser-based application for early-stage screening ahead of its launch. Users upload a STEP model, define the casting conditions and generate a tetrahedral mesh. At present, the model supports steels and cast irons. After mesh generation, a graph neural network calculates the predicted solidification time and local thermal modulus at every node. Engineers can examine the results within the casting by using section planes.
The images show a G20Mn5 pump casting with the risers already in place. The mesh comprises 12,373 nodes and 43,026 tetrahedra. In this example, the predicted solidification time is 423 seconds. The sectional view shows the pump body solidifying ahead of the upper and side risers. As intended, the final metal to solidify is located in these feeding elements, keeping liquid metal available as the casting contracts.
The hot-spot view highlights the area of the model that solidifies last. In this case, the highlighted areas are the risers, not the pump body. The local thermal modulus provides a complementary perspective, identifying the sections with greater thermal mass. Considered together, these results can help with the initial selection of risers, exothermic inserts and iron chills.
The purpose of FoundryFlash is to provide a rapid assessment of solidification, rather than final process approval. It therefore does not replace calculations for mould filling, porosity, stress or distortion.
Miknevič is currently working on an ultra-fast thermal solver based on the GPU-accelerated Finite Volume Method (FVM). Instead of depending on external proprietary simulation software, the solver uses validated physical datasets for commonly used metals, alloys and auxiliary materials as an automated “data factory.” This high-throughput pipeline generates synthetic data to train and refine the GNN model. The aim is not only to substantially improve prediction accuracy, but also to enable FoundryFlash to design and optimise casting feeding and rigging systems autonomously. Miknevič is currently discussing validation cases with several renowned foundries.