Physics-Led · Multi-Sensor · Bayesian Inference · Industrial Grade
Closing the
Yield Gap in
Mixed Scrap Streams
15,000 data points/sec · <20ms end-to-end inference · per-particle economic classification
Stochastic mass-balance control for secondary metal processing. Every particle on your conveyor belt classified by composition, density, and economic value — without stopping the line.
Yield Loss Is Happening on Every Shift
The loss is not visible — it leaves as undervalued mixed fraction. ρ-MATRIQS™ converts invisible loss into a traceable, recoverable number.
Multi-Sensor Fusion vs Single-Signal Classification
Single sensors produce unacceptable error rates on real industrial scrap. Error rate is bounded within ±2% only when all seven signals are fused and cross-validated.
❌ Single-Sensor Systems
- ✗Surface oxidation causes misclassification — sub-surface composition unknown
- ✗Binary output only — pass or reject, no alloy-grade resolution
- ✗No confidence scoring — classification errors undetected until audit
- ✗Manual recalibration required between material changes
✓ ρ-MATRIQS™ 7-Signal Fusion
- ✓7 signals cross-validate — errors self-cancel. Error rate bounded within ±2%
- ✓Alloy-grade resolution: Cu, Cu-Zn, Cu-Sn classified as separate fractions
- ✓Confidence score on every decision — uncertainty always explicit and logged
- ✓Self-calibrating — adapts to feed composition changes within 24h cycle
Not a sorter — a post-separation value recovery layer. Installs downstream of your existing ECS and magnet, on the belt you already have.
Check My Fit →Seven Ways to Read a Single Particle
Each sensor reads a different physical property. Alone, each can be defeated by surface state, particle geometry, or contamination. Combined, their cross-validated output produces a classification whose error rate is bounded within ±2% under mixed-material real-world conditions.
X-ray penetration up to 120mm particle depth. Denser materials attenuate more flux. Identifies true internal composition — unaffected by surface oxidation, paint, or coating.
1–2ms pulse cycle. Micro-plasma flash reveals elemental spectrum. Resolves Cu from Cu-Zn from Cu-Sn — grade-level, not class-level. No surface prep required.
Spectral range 400–2500nm. NIR fingerprints mineral types; VIS captures shape, colour, and surface texture at 4MP @ 120fps. Dual-mode cross-check eliminates single-band ambiguity.
Multi-frequency induction field separates Al from Cu, brass from zinc — by conductivity signature at each frequency. Resolves alloys that look identical to NIR.
First-pass ferrous gate. Instant identification of iron content — removes Fe contamination from the classification pipeline before non-ferrous stages run.
±1.5mm volumetric accuracy. Measures exact particle volume in milliseconds. Combined with XRT mass, yields true bulk density — not estimable from surface geometry alone.
Corrects all density readings for moisture content, temperature drift (±0.3°C resolution), vibration, and void space. Outputs real composition signal — not environmental noise.
One Installed System. One Certain Answer.
Each sensor's reading is weighted by its reliability for that material type. The Bayesian engine cross-checks, resolves conflicts, and produces a single classification with an explicit confidence score. Low-confidence particles pass through — they are never silently contaminated into a high-value fraction.
Measure → Decide → Recover
Each stage is physics-led, hardware-synchronised, and logged. Uncertainty is always explicit — never silently discarded.
Measure — 7 Signals, Simultaneously
Density, elemental composition, shape, conductivity, magnetic response — all read in under 20ms per particle. Belt encoder position-syncs every signal. No sequential sampling — all modalities fire in parallel.
Decide — Bayesian Inference Engine
Each signal is weighted by material-type reliability. Posterior probability P(material|signal) computed for all candidate classes. Mass-balance constraint applied per shift. Confidence score output on every decision.
Recover — Route + Record
Air jet fires at belt-encoder-tracked position. Grade certificate issued. CPCB/EPR compliance record created. All three outputs — physical sort, certificate, log — generated in the same decision cycle.
Measured on Deployed Installations
Where Conventional Sorting Falls Short
Optical and ECS systems perform well on clean, single-material streams. ρ-MATRIQS™ addresses the classification problem on mixed, variable, real-world material where single-signal systems produce unacceptable error rates.
| Capability | Optical / NIR | ECS | ρ-MATRIQS™ |
|---|---|---|---|
| Alloy-level identification (Cu vs Cu-Zn vs Cu-Sn) | ✗ Visual only — surface colour only | ✗ Conductivity only | ✓ LIBS + XRT composition fusion |
| Dirty / oxidised / real-world scrap | ✗ Fails on oxidised surfaces | ✓ Works through oxidation | ✓ Robust across contamination levels |
| Multi-metal in single pass | ⚠ Single output stream | ✗ Binary only — pass or reject | ✓ Cu / Al / Zn / reject simultaneously |
| 2–150mm particle range | ⚠ Narrow — clean, uniform pieces only | ⚠ Limited to specific size bands | ✓ Full shredder output range |
| Confidence score per decision | ✗ No — fixed threshold only | ✗ No — hardware separation only | ✓ 0–100% score, operator-configurable |
| Compliance traceability | ✗ None — no data recorded | ✗ None — no data recorded | ✓ Auto-recorded per sort decision |
| Installs on existing belt, no shutdown | ✗ New system, civil works required | ✗ New system, process change | ✓ Retrofit — mounts above existing belt |
Not sure which applies to your line? Send throughput and material type. Fit or no-fit answer before any demo is scheduled.
Check My Fit →The Complete 6-Layer Processing Stack
From physical material chaos to sorted, value-captured output. Each layer is defined, specified, and failure-handled. No layer hides errors from the next.
Mixed scrap stream with size/density/composition variability. Chaotic particle distribution on belt. This is where the yield gap originates — uncontrolled distribution creates sorting error before any sensor is active.
Seven sensing modalities sample every particle at belt speed. All sensor data is timestamped and position-synced via belt encoder. Raw signals are pre-processed at the edge node — no cloud upload.
Raw sensor signals are transformed into feature vectors per particle. The critical missing layer in most AI sorting systems — without it, classification is pattern-matching on noise.
Bayesian classification with mass-balance constraint. The model maintains Σ(input) ≈ Σ(classified output) — preventing grade drift over a shift. Output per particle: class label + confidence score + economic value weight.
Decision signal fires to air jet or mechanical diverter. Actuation only triggers when P(material) > configurable threshold AND economic value exceeds cut-off. Belt encoder drives position tracking — jet fires at the exact coordinate.
Sorted output is weighed and composition-sampled. Recovery % is computed per shift. Delta vs expected is fed back to recalibrate classification thresholds. Every 24h cycle improves system accuracy. Drift correction is automatic.
Feature Extraction — The Missing Layer
Most AI sorting systems go Sensor → AI → Output. This is why they produce high error rates on real-world oxidised scrap. Feature extraction converts raw sensor signals into physically-grounded, noise-corrected feature vectors before classification runs.
Hyperspectral data decomposed via PCA into material-specific absorption/reflection bands. Noise-corrected before classification. Each particle gets a 128-dimensional fingerprint cross-matched against our calibrated material library.
Acoustic impact sensing combined with TOF volumetric estimation yields density proxy without direct weighing. Separates ferrous fines from dense non-ferrous chunks before Bayesian classification runs.
Edge detection + surface texture analysis from RGB data. Differentiates wire/cable from sheet/chunk morphology. Form factor correlates with alloy grade in ASR streams — this layer captures that signal.
Spectral bands at 1400–1700nm detect oxide layers and surface contamination. Classification model compensates — sort decision is based on sub-surface material signature, not surface state.
Stochastic Inference Engine
Bayesian classification with mass-balance constraint. Every output has an explicit confidence score. Low-confidence particles pass through — they are never contaminated into high-value fractions.
Every sort session maintains Σ(input) ≈ Σ(classified output). If grade drift is detected across a shift — the model recalibrates priors automatically. This prevents systematic under/over-sorting of any specific material class.
Default confidence threshold: 0.78. Configurable per material class. When confidence is below threshold, particle passes through — not ejected. False positive rate is bounded; fraction contamination is controlled.
Every Failure Mode Is Handled
A system that only documents what happens when everything works is not an industrial system. The following failure modes are specified, tested, and handled at the hardware and software layer.
| Failure Mode | System Response | Belt Stops? |
|---|---|---|
|
Sensor dropout Handled — degraded mode |
Fallback weighting — remaining sensors redistribute confidence. No blind classification. | Never |
|
Low-confidence read Handled — no belt stop |
Particle passes through unejected. Zero contamination risk. Error rate bounded within ±2%. | Never |
|
Air jet failure Handled — degraded mode |
Automatic bypass mode. Belt continues. Fault logged and flagged to HMI within 200ms. | Never |
|
Encoder mismatch Handled — no belt stop |
Sync recalibration in <200ms. Position tracking restored without belt stop. | Never |
|
Edge compute fault Handled — no belt stop |
Redundant node takeover. Watchdog failover <200ms. Belt-stop never triggered by AI state. | Never |
|
Network/OTA loss Handled — no belt stop |
Last-known-good model held on-device. System continues fully air-gapped. Resync on reconnect. | Never |
Closed-Loop Yield Optimisation
Most sorting systems stop at actuation. ρ-MATRIQS™ measures the output, computes the recovery delta, and recalibrates the model — every 24 hours, automatically.
Air jet fires. Particle routed to fraction bin.
Fraction bins weighed at shift end.
Sorted recovery vs input mass computed.
Compare to model prediction.
Threshold + prior adjustment via OTA.
Improved accuracy from cycle N+1.
Deployment Spec, Latency Breakdown & Data Output Schema
What a plant engineer needs before a site visit. These are not marketing estimates — they are the parameters your DCS team will ask for.
Designed to Fit Your Enterprise Stack
Three-layer intelligence architecture — edge, plant, enterprise. Connects to existing DCS, SCADA, ERP via standard industrial protocols. No rip-and-replace. No PLC rewrite.
Physics-Led — Not a Black Box
Every decision is traceable to underlying physics equations. Not statistical pattern-matching on opaque training data — measurable physical properties, corrected for real-world conditions, combined with quantified uncertainty.
Physics-Based Density Modelling
Material density corrected for moisture, temperature drift (±0.3°C), vibration, and void space in real time. The output reflects actual material composition — not environmental signal.
Probabilistic Sensor Fusion
Each sensor's reading is weighted by its reliability for that material type. Uncertain or conflicting readings are detected, flagged, and excluded from the classification decision.
Uncertainty-Aware Decisions
Every sort decision carries a confidence score. Particles below threshold pass through — not ejected. Error rate bounded within ±2% under mixed-material real-world conditions.
Metallurgical Process Models
Industry-specific models for slag chemistry, alloy compositions, ore mineralogy, and recycled material grades. Decisions are calibrated for your downstream process — not lab defaults.
Real-Time Optimisation
Continuously optimises for your configured objective — maximum recovery, a target output grade, minimum energy per tonne, or lowest CO₂ intensity. Configurable per campaign.
IP-Protected Implementation
Algorithms, calibration parameters, and sensor orchestration logic protected under NDA. The science is documented here. The performance advantage is in the implementation depth.
Want the technical depth? Full sensor specifications, fusion methodology, and calibration protocols available on request under NDA.
Request NDA Access →Built for the Harshest Environments
Shredder halls. Open-air yards. Smelter adjacencies. ρ-MATRIQS™ is specified for conditions where consumer-grade equipment fails within weeks.
If ρ-MATRIQS™ goes offline for any reason — power fault, sensor error, network loss — your belt keeps running. The system fails to pass-through, not to stop. Zero production hours lost due to ρ-MATRIQS™ system state.
IP69K Rated
High-pressure washdown, dust, steam, chemical exposure. Continuous operation in smelters, shredder halls, and open-air yards. No additional protective enclosure required.
Temperature Compensated
All sensor readings corrected for temperature, moisture, and vibration automatically. ±0.3°C resolution. Accurate across full shift — even as plant environment changes.
Self-Calibrating
System monitors its own drift and recalibrates in background. No production stop for calibration runs. No manual tuning between shifts or material changes.
Plug-and-Play DCS Integration
Connects to existing DCS, SCADA, or PLC via OPC-UA, Modbus TCP, or Profinet. Operational within 72 hours of installation. No PLC rewrite required.
Operating Range: −10°C to 85°C
Calibrated for Indian industrial conditions — from cold startup to peak summer smelter adjacency. No warm-up period. No performance degradation at temperature extremes.
OTA Model Updates
AI model improvements delivered over-the-air during scheduled maintenance windows. Accuracy improves over time as system accumulates data from your specific material feed.
Every Classification
Is a Compliance Record
No separate data entry. No year-end audit scramble. No consultant to compile your CPCB return. Records are generated automatically on every particle decision.
- ✓CPCB Annual Return data — auto-generated from daily logs
- ✓EPR certificate per sorted material lot, auto-issued
- ✓Full material traceability — intake weight to output grade
- ✓CO₂ tracking per stream, per tonne, per day
Download the Integration Schema
Full integration schema including PLC interface spec, I/O mapping, mounting requirements, and edge compute footprint. Available immediately on request — no demo required to access the engineering documentation.
Industrial deployment ready · Built for Indian conditions · DPIIT Recognised