Metallurgy
Melt chemistry, inoculation and fade, sand and core behaviour, gating and solidification — the physics the agents have to respect.
Notes fromthe pour floor.
Engineering, metallurgy, safety and industry writing from the team building the autonomous operations layer for metal casting. Long, specific and occasionally uncomfortable.
[PLACEHOLDER] Article summaries are illustrative of Smelteon's editorial direction.
9Articles
6Topics
WeeklyCadence
Why soundness hasto be predictedbefore the pour.
Everything in a foundry is inspected after the value has been added. By the time an X-ray finds shrinkage porosity in a structural casting, the plant has already spent the melt energy, the mold, the pour, the cooling time and often the machining. We argue the only economically sane place to decide a casting's fate is before metal enters the mold — and lay out what a plant needs to get there: live melt and mold state, a fill and solidification model fast enough to run with the line, and a control path that can act on the answer.
- Inspection is a verdict, not a control input
- A twin that runs offline cannot change today's pour
- Prediction is only useful if something can act on it
- Attribution turns every grading into a training label
Everything wehave published.
Engineering · 9 min read
Why soundness has to be predicted before the pour
Inspection tells you what happened after the value was added. We argue the only economically sane place to decide a casting's fate is before metal enters the mold — and what it takes to get there.
Industry · 7 min read
The retiring pourer problem
Melting, molding and pouring are craft skills held in individual heads. What happens to a plant's process capability when three people retire in the same year — and what a model can and cannot inherit.
Safety · 11 min read
Bounded write-back: how an agent earns a setpoint
Shadow, assist, bounded. The three gates between a model and a furnace, why each exists, and what has to be true before a plant should let software move anything.
Quality · 8 min read
Attribution beats detection
Finding porosity is table stakes. Attributing it to a melt, mold, gating, pour or cooling condition is what turns an inspection result into a process change — and a training label.
Simulation · 10 min read
What a foundry digital twin has to do to be useful
Offline DOE simulation is a design-time tool. A twin that runs with the line has different requirements: speed, live state, and a promotion path into the cell plan.
Architecture · 9 min read
Federated learning without giving away the recipe
How defect and process-window priors can transfer between plants while alloy recipes, gating designs and customer programs never leave a tenant boundary.
Industry · 6 min read
Megacastings concentrate risk into a single pour
EV structural castings put an enormous amount of value into one fill event. Why that changes the economics of prediction, and why sampling inspection stops being sufficient.
Quality · 7 min read
The audit log is the product
Autonomy in a regulated manufacturing process is only as good as the evidence it produces. Designing an assurance-grade action log that an IATF 16949 auditor accepts.
Operations · 8 min read
Measuring melt energy per tonne, honestly
Energy is one of the largest controllable costs in a foundry and one of the worst-measured. A practical method for baselining it at the heat level before claiming any improvement.
What we writeabout.
Physical AI
Edge perception, control loops, model routing and what changes when your model's output moves a machine.
Safety & assurance
Autonomy gates, process windows, rollback design and the evidence model that makes it defensible.
Industry
Labour, lightweighting, megacastings, reshoring and the economics of a modern foundry.
“If your AI cannot explain a casting failure to the metallurgist who poured it, it is not ready to change the pour.”
We publishspecifics ornothing.
There is enough vague AI-in-manufacturing content in the world. Everything we publish should be usable by a process engineer: real parameters, real failure modes, real measurement methods, and honest labels on anything aspirational.
- Named parameters, not 'AI-powered optimisation'
- Measurement method stated before any claimed improvement
- Aspirational and placeholder content explicitly marked
- No customer numbers published without written approval
Three runningseries.
Series 01
The closed loop
- Perceive, plan, act, verify, log, retrain
- One post per stage of the loop
- Written for controls and process engineers
Series 02
Defect field guide
- Gas and shrinkage porosity, inclusions, cold shuts
- Misruns, hot tears, sand and core defects
- Cause, signal and control response for each
Series 03
Autonomy in practice
- Shadow-mode baselining
- Golden-set construction from plant ground truth
- Safety review and window authoring
9
Articles published
8 min
Median read time
3
Running series
0
Ghostwritten posts
Foundationalreading.
Safety · 11 min read
Bounded write-back: how an agent earns a setpoint
The three gates between a model and a furnace, and what has to be true before a plant lets software move anything.
Quality · 7 min read
The audit log is the product
Designing an assurance-grade action log that an IATF 16949 auditor accepts rather than tolerates.
Architecture · 9 min read
Federated learning without giving away the recipe
Transferring defect and process-window priors between plants while alloy IP never leaves a tenant boundary.
Housekeeping.
From foundry practitioners, yes. From content agencies, no. Write to hello@smelteon.com with the specific thing you want to explain.
Please do, with attribution and a link. If you are writing about the closed-loop model or the autonomy gates and want a technical review, we will read your draft.
There is no newsletter yet. Follow the company on LinkedIn or write to hello@smelteon.com and we will let you know when there is one.
Want the longversion in person?
Most of what we write started as an argument on a plant floor. Bring us yours.