During quotation and early design for a metalcasting program, I often need to check the intended direction of solidification before a complete simulation study is justified. Will the casting body freeze before the risers? Does a heavy junction need a different feeding approach? A conventional casting simulation is still the right tool for detailed process design, but the first review often calls for a quicker answer.

I am developing FoundryFlash for this early screening stage. The program accepts a STEP model and the selected casting conditions. Its current material range covers steels and cast irons. After meshing, it calculates solidification time and local thermal modulus at the mesh nodes. The engineer can inspect both fields in a browser, use section planes to look inside the part, and read a value at the cursor.

The calculations are produced by a graph-neural-network surrogate. This replaces hours of reference simulation on a multicore processor with a result that is available in seconds after the mesh is ready. I use that speed to compare concepts, while keeping full simulation as the route to process approval.

FoundryFlash is still a prelaunch application, and I am working on the accuracy of the model.

Solidification time and thermal modulus

The solidification-time field shows when each node crosses the solidus. A surface view alone can hide the thermal behavior inside a heavy section, so the browser includes movable cutting planes and a still-liquid display. The latter shows which part of the metal has not yet crossed the solidus at a selected time.

For a feeding layout, the desired result is not simply the shortest possible time. The sequence should end in the risers, where liquid metal is available to compensate for contraction in the casting.

The second field is local thermal modulus, calculated at the nodes rather than reported only as one V/A value for the complete casting. It gives the engineer a geometric measure of the sections that are more difficult to feed.

Used together, solidification time and local modulus can support the initial selection of risers and exothermic sleeves or inserts. They also can help identify where an iron chill may be needed to change the freezing sequence.

I use these fields as engineering aids. They do not design a feeding system automatically. The modulus field cannot judge whether a proposed riser neck is practical, and solidification time alone cannot predict every shrinkage defect. The engineer still has to interpret the freezing order and section geometry before moving to the complete process model.

What the pump casting shows

The G20Mn5 pump casting in the accompanying figures already includes its risers and was screened with a no-bake furan mold. The browser gives a predicted total solidification time of 423 seconds. The sectioned view shows the casting body freezing first. Metal remains liquid longest in the attached risers, which is the intended result because the risers must continue feeding the casting as it contracts.

The upper cylindrical feature and the tall side feature are therefore not defect locations in this example. They are feeding elements. Their later solidification provides a useful first check that the sequence runs away from the casting and into the risers.

The local-modulus view also shows why these regions remain liquid: their effective section is heavier than most of the pump housing.

For a numerical check, I ran a FEniCSx reference and the frozen prediction on the same 12,373-node mesh, with no interpolation between them. The reference reached 498.1 seconds and the prediction 422.8 seconds. Mean absolute error was 44.7 seconds.

The model reproduces the important order of events, with the risers freezing after the casting body, but it understates the final time by 15.1%. That error keeps absolute time out of process approval and is one reason I am improving the model.

Current development work

I am now training the model on real production casting geometries that include exothermic inserts and iron chills. An exothermic insert delays freezing in the riser area, while a chill removes heat locally and accelerates solidification.

The model must reproduce both effects without losing the geometry response it already provides. I am checking the resulting fields on unfamiliar castings as this work progresses.

FoundryFlash can screen steels and cast irons, calculate nodal solidification time and local thermal modulus, and help an engineer make an initial choice of feeding aids. It does not calculate mold filling, melt flow, porosity, stress or distortion. A full simulation is still required before a feeding system is approved or a casting process is signed off.

The present numerical comparisons use my FEniCSx reference, so they do not yet constitute independent industrial validation. The next step will be comparison with real castings, including measured cooling behavior and observed shrinkage. I also plan to validate the predictions against simulations in specialist casting packages such as MAGMASOFT and ProCAST.

I am currently discussing validation projects and access to suitable cases with foundry companies. Those discussions are still in progress, so I do not present commercial-solver or shop-floor results here.

In a pilot I intend to check whether the casting freezes toward the risers and whether sleeves and chills act in the intended places. Then, predicted times will be compared with measured or independently simulated values.