Adaptive charge and chemistry
Chemistry targets held across scrap-charge variation, alloy changes and furnace state — with alloy additions and corrections proposed per heat, not per recipe book.
Everything theloop can holdon target.
Feature by feature, the controls, sensors, models and guardrails that make a foundry run closed-loop — from the charge going into the furnace to the finished casting leaving the fettling cell.
9Agent domains
12+Connector families
100%Actions logged
Built with foundries and die casters melting iron, steel, aluminium, magnesium and copper
Ironhaus GroupMeridian CastingsVulcan DuctileAurora Light MetalsKestrel Die CastingNorthfield Foundry Co.
[PLACEHOLDER] Design-partner names are illustrative until first references are signed.
Chemistry that holdsheat after heat.
Melting is the plant's biggest energy line and its first quality gate. Meltra treats it as a control problem, not a recipe.
Temperature and tap control
Superheat, holding and tap temperature controlled against the pour plan, so metal arrives at the mold at the temperature the twin assumed.
Degassing and inoculation
Degassing, inoculation and nodularization timing controlled and verified with thermal analysis, so fade never silently changes the metallurgy.
Melt energy optimization
Furnace energy windows scheduled with cuOpt against tariff, melt readiness and mold readiness — melt loss and kWh per tonne both fall.
Geometry decided early,not discovered late.
Coreon controls the sand system and the core cell so mold quality is a controlled variable instead of an inspection finding.
Sand property control
Compactability, moisture, green strength and return-sand condition tracked and corrected before a bad batch reaches the molding line.
Core strength and geometry
Shot parameters, cure and core strength controlled per core box, with dimensional verification tied back to the part.
Mold quality scoring
Every mold scored against the twin's expectation before metal is committed to it — no pouring good metal into a bad mold.
Dimensional conformance
Dimensional faults linked to sand, core and mold conditions instead of being written off as line variation.
The two steps thatdecide soundness.
01Stream, tilt, temperature, timing
Pourbot controls ladle attitude, stream shape, pour temperature and fill time per geometry — the difference between a filled mold and a misrun or cold shut, executed the same way on every mold.
- Pour vision on the stream, not just the ladle
- Fill time held per part and per mold state
- Turbulence and oxide entrainment minimized
- Fill
- Freeze
- Predict
Every pore, inclusionand cold shut.
Poreon fuses X-ray, CT, vision and dimensional data into one verdict per casting — and one root cause for the agent upstream.
X-ray and CT fusion
DICONDE image and volume inference on Jetson and IGX, with CT batch triage inside minutes per part family rather than a sampling plan.
Surface and vision defects
Sand defects, cold shuts, misruns and surface faults detected at line rate across 8–24 camera or detector streams per line.
Dimensional and microstructure
Dimensional scans and microstructure indicators graded against the part spec and the twin's as-cast prediction.
Root cause, sent upstream
Every defect is attributed to melt, sand, gating, pour or cooling conditions and returned to the responsible agent as a training label.
Nine defect families,one grading model.
The defect taxonomy the platform senses, predicts and attributes — the vocabulary a foundry actually argues in.
Gas porosity
Entrapped or dissolved gas, predicted from melt hydrogen state, degassing history, pour turbulence and mold permeability.
Shrinkage porosity
Feeding failure during solidification, predicted from cooling curves, gating and riser performance in the twin.
Inclusions
Oxides, slag and dross, attributed to melt handling, ladle practice, stream turbulence and filtration.
Cold shuts
Incomplete fusion of two metal fronts, driven by pour temperature, fill time and gating geometry.
Misruns
Incomplete fill, driven by pour temperature, fluidity, fill rate and mold venting.
Hot tears
Tearing during solidification contraction, predicted from thermal gradients, mold restraint and geometry features.
Sand defects
Inclusions, scabs, erosion and penetration linked to sand properties, compaction and mold handling.
Dimensional faults
Deviation from print, linked to core position, mold conditions, shrinkage and shakeout timing.
Microstructure drift
Nodularity, graphite morphology and phase drift, tied to chemistry, inoculation fade and cooling rate.
The rest of the plantis in the loop too.
Robotic fettling and grinding
Cut-off, grinding and deburring paths learned per geometry and executed hands-free, removing the hottest, dustiest, most injury-prone manual work in the plant.
Robot and casting handling
Pouring robots, mold handling and casting handling commanded inside safety-rated envelopes with Isaac-based motion and full traceability.
Yield, scrap and takt
Yield, melt loss, scrap and moving-line takt balanced across melt, mold and pour, with sub-60-second rescheduling when the line is disrupted.
Genealogy and conformance
Right-first-time, non-conformance and full casting genealogy maintained for IATF 16949, ISO 9001, AS9100 and NADCAP evidence.
$ smelteon cell status P-04cell P-04 · pour · autonomy=bounded · window=OKpour temp 1398 °C (window 1385–1410)fill time 6.8 s (target 6.5–7.2)twin plan castwin:solidification-v7 (approved)$ smelteon predict --heat 4412 --part crankshaft-4cyl-rev-cshrinkage porosity @ journal 3 ....... 2.1% riskcold shut ............................ <0.5% riskrecommendation: +8 °C pour temp, riser 2 open 0.4 s laterstatus: within approved window → awaiting approver$ smelteon audit tail --cell P-04 --limit 216:04:11 plan applied · approver=metallurgist-on-shift16:09:52 outcome graded · poreon: sound · genealogy written
Operable by thepeople on shift.
The CLI and console speak in foundry terms — heats, molds, pours, windows, approvals — not in model names.
30–120
FPS vision inference per line
< 2 min
CT batch triage per part family
< 60s
Line reschedule after disruption
12–40
Models served per factory
[ASPIRATIONAL] Performance targets from Smelteon's GPU utilization plan.
The things engineersask second.
That is the point. Predictions come from melt, sand, mold, pour and cooling signals, and Poreon's X-ray and CT verdicts are the labels that keep them honest. Inspection stays the ground truth; prediction is what lets you act before the value is added.
No. Melt chemistry, sand and mold control, solidification control and inspection all run on machines you already have. Robotic pouring and fettling are where robot integration matters, and those are separate wedges.
Castwin simulates the new geometry and alloy against the plant's own history, proposes gating and process windows, and the models start in shadow mode on that part until accuracy clears the golden-set gate.
Pick the featurethat fixes yourworst number.
Bring us the cell where scrap, rework, melt loss or inspection backlog costs the most. We will baseline it and show what the loop moves.