Proof of concept · real-time 3D mold flow

Mold Flow Studio

Injection molding pushes molten plastic into a steel mold. Before a mold is cut, engineers simulate the fill to catch parts that come out incomplete, weak or slow to make. Those simulations take minutes to hours. Here a neural network predicts the fill in under a second, and a physics simulation runs right next to it so you can see whether the AI got it right.

How it works

1
Pick a partChoose an example, build your own shape, or open a STEP file.
2
Set the gate and machineClick the part to move the gate, the point where plastic enters. Drag the temperature and pressure sliders.
3
Compare AI with physicsThe AI answers first on the left. The slower physics check fills in on the right.

1Part

Change the shape

size, walls, ribs, holes, slots

Every example is built with these controls. Change anything and the AI answers right away, even for shapes it has never seen.

Outline

Length

Width

Wall height

Floor thickness

Wall thickness

Floor ribs

Round holes

Vent slots

Thick top rim Screw bosses Thin wall panel Thin floor logo

Open a STEP file

experimental

Simple open boxes, trays and lids only: a flat floor with straight walls, up to 95 × 85 mm and 20 mm deep. The file stays in your browser.

2Machine settings

Melt temperatureHotter plastic is runnier and stays liquid longer.

Mold temperatureA warmer mold freezes the plastic more slowly.

Injection pressureMore pressure pushes the plastic in faster.

Part–

AI prediction–

Physics check–

Do they agree?–

AI prediction–

What the neural network expects. It learned from 12,000 earlier simulations and answers in under a second.

Physics simulation–

The step-by-step flow calculation the AI was trained to imitate. Slower, and the reference for checking the AI.

–

Weld lines Air traps (physics) Run physics after every change

Drag to turn the part. Hover a spot to read its values. Click the floor, a wall or the rim to move the gate ★ there.

3What to watch for

AI vs physics, side by side

	AI	Physics

4Find the best gate

The AI ranks every tested gate position at the current settings by fill completeness, fill time and weld lines. Physics then checks its top 20 and keeps the best.

Which settings work?

The AI tries 42 combinations of melt temperature (across, °C) and pressure (down, MPa) for this gate. Numbers are fill times in seconds. Red squares leave part of the mold empty, with the empty share shown. Click a square to use it.

How good is the AI?

Tested on 100 parts and settings it never saw during training, against the physics simulation.

–typical error in fill time

–of short-shot areas caught (spots that never fill)

–for one AI prediction in your browser

–for the same fill with physics in your browser

On real 3D parts, full commercial simulations take 30 to 60 minutes, so the gap grows with part size. That speed is what makes it practical to try every gate and every setting.

How it works

Done once, before you opened this page

Simulate
Train
ShipThe trained network, about 1.4 MB of numbers, is built into this page.

Happens live, every time you change something

Describe the partThe surface is cut into 2.5 mm square patches. Each patch knows its wall thickness, its neighbours and its distance from the gate.
AI predictsThe network passes information between neighbouring patches 12 times, then outputs fill time, fill or no-fill and pressure for every patch. Frozen skin, clamp force and weld lines are worked out from that.
Physics checksThe simulator steps the melt front forward patch by patch, solving for pressure each time. Air traps come from this result only.
CompareBoth results are drawn side by side, and the strip above the views says how closely they agree.

Terms used on this page

Gate ★: Where molten plastic enters the mold. Its position changes everything about the fill.
Fill time: When the melt reaches each spot, from 0 s at the gate.
Short shot: Plastic freezes before reaching some areas, so the part comes out incomplete.
Weld line: Where two melt fronts meet. Often a visible line and a weak spot.
Air trap: A pocket the melt surrounds with no edge for the air to escape. Can cause burns or voids.
Frozen skin: Plastic freezes inward from the cold mold walls. The more of a wall that has frozen, the harder it is for melt to flow through.
Clamp force: How hard the machine must hold the mold shut against the melt pressure. Sets the machine size.
Patch: One 2.5 mm square of the part's surface. The AI and the physics both work patch by patch.

What this prototype simplifies

One family of shapes: a flat floor of any outline, with holes and slots, and straight side walls folded up along its outer edge. No draft angles, fillets, curved walls, standing bosses or snap fits. Up to 95 × 85 mm and 20 mm deep.
The surface is cut into 2.5 mm square patches and melt flows along the wall. Flow through the wall's thickness and true 3D effects at corners and thick sections are not modelled.
The simulator uses Hele-Shaw flow with a fixed injection pressure and a frozen skin that grows with the square root of time. It has no packing, cooling or warpage stage, and it is not validated against Moldflow or a real molded part.
Process settings use simple rules for one generic plastic: viscosity falls about 2.3% per °C of melt temperature, and the frozen skin grows faster with a cold mold or a melt close to its no-flow temperature. They are not fitted to real ABS or PP data. Temperature itself is not simulated.
"Frozen skin" is the share of each wall frozen by the end of fill, worked out from fill times with the simulator's skin formula. Clamp force uses end-of-fill pressure on the floor's projected area. Real machines need more, because packing pressure is higher.
The gate optimizer uses a simple score we picked by hand (short shot first, then fill time, then weld lines). It is not a validated measure of part quality. The AI alone often ranks the true best gate in its top 20 but not first, which is why physics confirms the shortlist.
STEP import is experimental and only recognises simple open boxes, trays and lids.