Industries / Mining and materials / Plastics and rubber
Injection molding
Injection molding melts plastic pellets and forces the melt under pressure into a closed, cooled mold, where it hardens into a part.
7
Milestones achieved, from 1872 to 2016.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1872
Plunger injection patented
US patent 133,229 describes a plunger in a heated cylinder that forces celluloid through a nozzle into a mold. SPE's history calls it the patent for the injection molding machine, and Tadmor and Gogos's textbook says Hyatt "invented the thermoplastics injection-molding machine in 1872".
Credit
John Wesley Hyatt and Isaiah Smith Hyatt, Albany, New York
1946
A screw melts the plastic
A turning screw inside the heated barrel melted and mixed the plastic before it was injected. SPE calls Hendry's the first screw injection molding machine and says almost all machines today use the technique.
LessonPlastic insulates against heat, so a barrel heated only from outside melts the middle slowly, and working the plastic with a screw solves that.
Credit
James W. Hendry
1952
The screw becomes the ram
In a reciprocating-screw machine, one screw turns to melt the plastic, then slides forward like a piston to inject it. The Plastics Hall of Fame says plunger machines were limited to shots of about 24 ounces (680 g), and this design opened the way to larger parts with thinner walls.
Credit
William H. Willert, Egan Machinery, New Jersey (filed 1952, granted 1956). SPE notes that some credit Hans Beck of BAST, Germany, with a 1943 design
1978
Simulating the fill
Software began to predict, from a drawing, the pressure needed to fill a mold before any steel was cut. By 1993 it also predicted warpage, which is how much a part bends after molding. It is now a routine tool: a 2024 university study used it to decide where to put sensors in a mold.
Credit
Colin Austin, who founded Moldflow in Australia in 1978
1983
All-electric machines
Electric servo motors replaced hydraulics to drive the machine's motions. Nissei says its MM-5 of 1983 was the first such machine, and Fanuc's own timeline dates its Autoshot to 1984. In 2025 the molder Evco Plastics said one large all-electric press used a quarter of the energy of its other presses of that size, by Evco's own figures.
nisseiplastic.comfanuc.co.jpplasticsmachinerymanufacturing.com
Credit
Nissei Plastic Industrial; Fanuc
1985
Filling and packing decoupled
Molders began to set filling the mold and packing it as separate stages. Packing means pushing in extra plastic as the part shrinks. They guide both with a sensor that measures pressure inside the mold cavity, instead of setting the machine by feel. A trade profile traces much of today's molding practice to this work.
Credit
Rod Groleau, founder of RJG
2016
One data format for machines
EUROMAP 77 gives machines from different makers one OPC UA interface. OPC UA is an open protocol that industrial machines use to share data, so each machine can report its settings and shot data to factory software. The first release came out before the K trade fair in October 2016, and SKZ and Fraunhofer IPA later used it to record a public dataset.
LessonThe older EUROMAP 63 standard swapped text files, which the OPC Foundation calls "not state of the art".
opcfoundation.orgreference.opcfoundation.orgb2share.eudat.eu
Credit
Arburg, Engel, Ferromatik Milacron, KraussMaffei, Negri Bossi, Netstal, Sumitomo (SHI) Demag and Wittmann Battenfeld, with the OPC Foundation
Watch for
A cell that sets itself
Engel says its molding cell sets its own machine settings from quality targets the operator enters. It says the cell can save up to 5% of material by running near the lower tolerance limit. Engel calls it the "world's first industrial solution for an autonomous, self-regulating injection moulding cell". It names no customer and no outside test.
Achieved whena molder running the cell in production, or an independent lab, publishes months of material and scrap figures against a normal cell.
Claimed by
Engel, September 2025 (shown at K 2025, Düsseldorf, October 2025)
AI assistant at the press
Chen Hsong's iChen AI Molder is software on the machine. It calculates first settings, suggests fixes for defects and adjusts settings during production. A trade report says the company designed it "to replace the junior operator" and that it is installed on about 300 presses. Both are company figures.
Achieved whena molder or journal not paid by Chen Hsong publishes setup shots, start-up scrap and time to first good part, with and without the assistant.
Claimed by
Chen Hsong, October 2025
Simulation 1,000 times faster
Simcon says its Cadmould AI Solver, built with Emmi AI, predicts how a mold fills "up to 1000 times faster than traditional numerical solvers". Simcon says it is for early design work and does not replace the classic solver.
Achieved whenindependent users publish its error against the classic solver and against molded parts, on parts it was not trained on.
Claimed by
Simcon, March 2026
Open
01
Few-shot setup in the open
Open-source software that sets up a new mold on a small lab injection machine. It proposes settings, reads each shot's part weight and photo, and reaches good parts in as few shots as possible. Run it against a trained human setter on the same molds.
The briefCredit · unclaimed02
Fixes tested on a press
A public set of at least 100 defect cases made on a lab machine. Each case holds a photo of the bad part, that shot's settings, the fix a skilled molder chose, and the measured result after the fix. AI assistants are then scored by whether their suggested fix works when it is tried on the machine.
The briefCredit · unclaimed
Open 01 · the brief
Few-shot setup in the open
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Open-source software that sets up a new mold on a small lab injection machine. It proposes settings, reads each shot's part weight and photo, and reaches good parts in as few shots as possible. Run it against a trained human setter on the same molds.
Why it is still open
Setting up a new mold is still done by trial or by planned rounds of experiments. A 2024 study at Atlantic Technological University in Ireland used Bayesian optimization, a method that picks each next trial from the results so far. It needed 16 experiments where a standard plan needed 31, by the authors' count. The authors found "few practical studies" of the method in manufacturing "and none in relation to polymer processing". They released no code. The machine makers' versions are closed and unmeasured by outsiders (see Watch for).
Built with what exists
A mini injection machine sold to test labs and universities, such as the Electro-Matic MiniJector Model 55E, run by trained lab staff; BoTorch, Meta's open Bayesian optimization library; a lab balance; a USB camera; calipers.
First step
Log one experienced setter setting up one mold, with the settings, part weight and a photo for every shot, for 50 shots, and publish the log.
Done looks like
On 3 molds and 2 plastics, the software reaches good parts in 15 shots or fewer in at least 8 of 10 runs. A good part is within 0.5% of a reference part's weight, has no short shot (a part that did not fill) and no flash (thin excess plastic at the mold's edges), and meets one drawing dimension. A human setter's shot counts on the same tasks are published alongside, with code, shot logs and photos.
Open 02 · the brief
Fixes tested on a press
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of at least 100 defect cases made on a lab machine. Each case holds a photo of the bad part, that shot's settings, the fix a skilled molder chose, and the measured result after the fix. AI assistants are then scored by whether their suggested fix works when it is tried on the machine.
Why it is still open
The closest test found, KAIST's IM-Chat (July 2025), had experts and GPT-4o grade language-model answers on 160 tasks and did not try any fix on a machine. A 2024 defect image set cites "the lack of publicly available injection-molded product defect datasets", and its files have not been uploaded.
Built with what exists
The same kind of lab machine, run by trained staff; a phone or USB camera; the SKZ and Fraunhofer IPA dataset (CC BY 4.0) as a model for the data layout; small open-weight AI models such as Liquid AI's LFM3, which Liquid says can be fine-tuned on one graphics chip (Note: "One GPU to fine-tune"); the assistants that machine makers offer.
First step
On one mold, make 20 defects on purpose by changing one setting at a time, such as lowering hold pressure to cause sink marks, and photograph and log each one.
Done looks like
100 published cases covering at least 5 defect types: short shot, flash, sink mark (a shallow dent over a thick section), burn and streaks. Each has before-and-after photos and settings. A table shows, for at least 3 assistants, how often the first suggested fix produced a good part.