Deep Tech • Circular Economy • Spectroscopy

Seeing the Invisible

The Physics, Economics, and Intelligence of Black Plastic Recovery

Every year, millions of tons of black plastics are systematically ejected from recycling streams and sent to landfill—not because they lack value, but because sensors literally cannot see them.

Veriprajna's solution shifts observation from Near-Infrared (blind) to Mid-Wave Infrared (chemistry-based), using 1D-Convolutional Neural Networks to interpret spectral signatures invisible to the human eye and standard AI.

📄 Read Full Whitepaper
9%
Global Plastic Recycling Rate
353M tons waste/year
3-15%
Black Plastic in Waste Stream
Invisible to NIR sensors
$1,200
Per Ton Recycled PP Value
Parity with virgin resin
<2mo
ROI Payback Period
Mid-sized MRF

The Black Hole in the Circular Economy

A systemic failure invisible to the human eye but catastrophic to the environment and the balance sheet alike.

The Carbon Black Singularity

Carbon Black is a paracrystalline form of carbon from incomplete petroleum combustion. It acts as a near-perfect black body absorber across visible/NIR spectrum (700-1700nm).

NIR Sensor → Reflectance: 0%
Signal: NULL → RESIDUE
💰

The Economic Magnitude

For a MRF processing 50,000 tons/year, 3.3% black plastic = 1,650 tons discarded. At $1,130/ton, that's $1.86M in lost revenue annually.

Automotive: 240kg plastic/vehicle
Electronics: Black ABS casings
FMCG: Black packaging trays
⚖️

The Regulatory Imperative

EU PPWR mandates all packaging must be recyclable by 2030. If sensors can't see it, it's legally classified as unrecyclable—an existential threat to brands.

Landfill tax: €100/ton (EU)
EPR laws: Producer liability
ETS: Carbon cost for incineration

"While a human operator can clearly see a black tray on a belt, the machine vision system effectively sees nothing. This is a failure of physics that no amount of computer vision contrast adjustment or prompt engineering can resolve. One cannot enhance a signal that was never captured."

— Veriprajna Technical Whitepaper, 2024

The Physics of Blindness

To engineer a solution, one must first dissect the failure mode of incumbent technology.

Standard NIR Spectroscopy

The recycling industry's workhorse sensor operates at 900-1700nm. For colored plastics, light penetrates, interacts with molecular bonds (C-H, N-H, O-H), and reflects back with a unique spectral fingerprint.

The Failure Mode:

When light strikes black plastic, carbon black pigment absorbs incident radiation before it can interact with the polymer matrix. Sensor receives effectively zero return signal.

Black object on black belt = zero contrast
Flatline signal → sorting computer sees "empty belt"
Pneumatic ejectors remain silent → valuable polymer to residue

Why "AI Wrappers" Cannot Fix This

We exist in a technological epoch defined by LLMs and Generative AI. Many "AI Consultancies" offer to wrap GPT-4/Claude via API to solve enterprise problems.

The Category Error:

Generative AI models require input data to function. In black plastic sorting via NIR, the input data is a null set. There is no spectral curve to analyze; there is only noise.

Input Void: No spectral data captured = nothing to classify
Hallucination Risk: Forcing LLM to classify noise = guessing, not sensing
Latency: Cloud API calls (100ms-seconds) vs required <10-20ms decisions

Veriprajna Asserts:

The solution lies in Deep Tech—the rigorous application of scientific discovery to engineering problems. We do not try to make the AI "smarter" at guessing; we change the physics of the sensor to make the invisible visible.

The Photonic Solution: Mid-Wave Infrared

If the data does not exist in the Near-Infrared, we must shift our gaze to a different window of the electromagnetic spectrum.

The Electromagnetic Window

Carbon black's absorption is not infinite. As wavelength increases, absorption decreases. By moving to MWIR (2.7-5.3 µm), we cross a threshold where pigment becomes sufficiently transparent.

NIR/SWIR (0.9-2.5 µm)

Dominated by "overtones" of molecular vibrations. Weak signals, easily swamped by carbon black absorption.

MWIR (2.7-5.3 µm)

Contains fundamental vibrations of polymer molecules. C-H stretching, C=O carbonyl stretches—orders of magnitude stronger signals.

Chemical Vision vs Computer Vision

Hyperspectral Imaging represents a paradigm shift from analyzing external appearance to analyzing intrinsic composition.

Computer Vision (RGB)

Analyzes shape, texture, label, color. Unreliable for crushed, dirty waste. Fragment of black bumper looks identical to black tray.

Hyperspectral Imaging (HSI)

Analyzes chemistry. Sees chemical signatures:

  • PE: C-H stretching @ 3.4 µm
  • PS: Aromatic C-H modes (sharp, unique)
  • ABS: Flame retardants detectable (WEEE separation)

Specim FX50: Industrial-Grade Hardware

Veriprajna builds upon the only commercially viable sensor covering the full 2.7-5.3 µm range. The engineering complexity highlights the "Deep Tech" nature of the solution.

154
Spectral Bands
High-res sampling
380
Frames/Second
Real-time speed
77K
Cryogenic Cooling
Stirling cooler
MCT
HgCdTe Detector
Exotic semiconductor
Unlike Silicon (RGB) or InGaAs (NIR) sensors, MWIR requires cryogenic cooling to prevent thermal noise—significantly increasing the barrier to entry.

Spectral Fingerprints: The Divergence

Materials identical in visible light have unique spectral fingerprints in MWIR. This chart shows reflectance across wavelengths—note the dramatic separation in the MWIR region.

Visible (400-700nm)

All black plastics overlap near 0%. Carbon black absorbs all visible photons—sensors receive no signal for classification.

MWIR (2700-5300nm)

Lines diverge significantly. Fundamental stretching/bending vibrations are orders of magnitude stronger than NIR overtones.

  • PS: Aromatic C-H stretching (unique sharp peaks)
  • PP: Aliphatic C-H absorptions (methyl groups)
  • PE: Strong C-H bands @ 3.4 µm

This capability turns sorting from a visual guessing game into quantitative chemical analysis performed at conveyor belt speed.

Reflectance Spectra: Black Plastics

Simulated based on fundamental vibrational spectroscopy. NIR region shows convergence; MWIR shows clear separation.

The Cognitive Engine: 1D-Convolutional Neural Networks

Capturing the MWIR hypercube is only half the battle. We need AI capable of interpreting complex chemical signatures with near-zero latency.

Why 2D-CNNs Fail in Waste

Standard AI image recognition (ResNet, YOLO) uses 2D-CNNs designed to recognize spatial features: the curve of a cat's ear, the rectangle of a license plate.

Spatial Irrelevance:

Plastic bottles are crushed, twisted, torn. Fragment of black bumper looks identical to black tray spatially.

Spectral Blindness:

Treating 154 spectral bands as "color channels" fails to capture subtle high-frequency relationships defining chemical fingerprints.

The 1D-CNN Paradigm

Veriprajna treats the problem not as image recognition, but as signal processing. For every pixel, we extract a 1D vector of 154 values—the spectrum.

Input:

1×154 vector of normalized reflectance values

Conv1D Layers:

Linear kernels learn spectral shapes: sharp drop @ 3.4 µm, broad shoulder @ 4.0 µm, doublet peaks

Output:

Softmax probability: [PP: 99%, PE: 0.5%, Other: 0.5%]

01

Hypercube (x,y,λ)

Specim FX50 generates 3D data: every pixel contains 154-band continuous spectrum for chemical analysis.

640×N×154 tensor
02

Conv1D Kernels

Linear kernels slide over spectrum, learning to detect specific spectral shapes—the "grammar" of molecular bonds.

Learn C-H, C=O peaks
03

Edge Inference

NVIDIA Jetson AGX Orin runs optimized model. No cloud latency. Deterministic <5ms inference time.

TensorRT optimized
04

Real-Time Actuation

PLC interface triggers pneumatic ejection or robotic arm within total <20ms budget at 3m/s belt speed.

Jitter = contamination

Why 1D-CNNs Over Transformers/LSTMs?

Academic Models: Too Slow

Belt speed: 3m/s. Camera→Nozzle distance: ~1m. Time budget: 300-500ms total (acquisition, preprocessing, inference, segmentation, valve timing).

  • Transformers: O(N²) complexity, quadratic with sequence length
  • LSTMs: Sequential, difficult to parallelize, high latency

1D-CNN: Industrial Speed

Jitter = Contamination. If air jet fires 50ms late, it hits the wrong object. 1D-CNNs provide deterministic, sub-5ms inference.

  • Linear complexity: O(N), highly parallelizable on GPU
  • TensorRT optimized: Thousands of frames/second on edge hardware
ROI Calculator

Calculate Your Recovery Potential

Adjust parameters based on your facility's throughput and composition to model economic impact

50,000 tons

Typical MRF range: 10K-200K tons/year

5%

Industry data: 3-15% of waste stream depending on region

90%

Veriprajna achieves 90%+ with MWIR+1D-CNN

Annual Revenue
$2.03M
@ $900/ton avg
Payback Period
1.8mo
$300K CAPEX

Case Study: Mid-Sized European MRF

Current State (NIR Blind)

Throughput: 50,000 tons/year
Black plastic (5%): 2,500 tons
Incineration cost: -€250,000
Lost material value: -€2.25M
Total Impact: -€2.5M/year

With Veriprajna MWIR

Recovery rate: 90% (2,250 tons)
Sales revenue: +€2,025,000
Disposal savings: +€225,000
System CAPEX: -€280,000
Annual Impact: +€2.25M
Payback Period: < 2 Months

Strategic Positioning: The "Deep Tech" Moat

In a marketplace crowded with "AI for X" startups, Veriprajna distinguishes itself through vertical depth and physical integration.

🧬

Proprietary Data

MWIR spectral signatures of dirty, crushed, real-world waste are not on the public internet. We generate and own this dataset through physical deployments.

Cannot be replicated by scraping or API calls
🔧

Physical Moat

Our solution requires integration of cryogenic hardware, edge computing, and robotics. This cannot be copied by developers in a coffee shop—it requires engineering rigor.

High barrier to entry: hardware + industrial access
🔄

Retrofit Capability

Many MRFs have good conveyor systems but outdated sensor heads. Veriprajna upgrades the "eyes and brain" without replacing the "body"—lowering CAPEX.

Integrate with Tomra, Steinert, existing infrastructure

Escaping the "Wrapper" Trap

❌ AI Wrapper Business Model

  • • Build UI on top of someone else's API (GPT-4, Claude)
  • • No moat: if OpenAI adds your feature, business evaporates
  • • Constrained by underlying model limitations
  • • Cannot "prompt" away physics problems

✓ Veriprajna Deep Tech Model

  • • Change the sensor physics to capture chemical reality
  • • Own proprietary datasets from industrial deployments
  • • Custom 1D-CNN architectures for spectral processing
  • • Vertical integration: hardware + AI + edge computing

Beyond Sorting: The Future

Once we can "see" chemistry, we unlock value beyond simple material recovery.

📋

Digital Product Passports

Track exactly what materials pass through the recycling chain. Provide granular data on polymer types, contamination levels, and throughput for EU DPP compliance.

Quality Assurance

Certify bale purity before shipping to converters. Real-time spectral analysis reduces rejected loads, increasing customer trust and premium pricing.

🏢

Brand EPR Audits

Provide producers with data on how much of their specific black packaging is actually recovered, helping meet Extended Producer Responsibility reporting requirements.

Stop Looking at the Picture. Start Looking at the Chemistry.

Veriprajna's MWIR Hyperspectral + 1D-CNN solution doesn't just improve recovery rates—it fundamentally changes the physics of observation.

Schedule a technical consultation to discuss your facility's black plastic recovery potential.

Deep Tech Consultation

  • • MWIR spectroscopy feasibility assessment
  • • Custom ROI modeling for your throughput
  • • 1D-CNN architecture review for your waste stream
  • • Integration roadmap with existing infrastructure

Pilot Deployment Program

  • • On-site Specim FX50 demonstration
  • • Real-time dashboard with spectral analysis
  • • Operator training & knowledge transfer
  • • Comprehensive performance report with chemistry data
WhatsApp: Deep Tech Discussion
📄 Read Full Technical Whitepaper (17 Pages)

Complete engineering report: MWIR photonics, 1D-CNN architectures, spectral unmixing mathematics, sensor fusion, edge computing implementation, economic analysis, regulatory alignment, comprehensive works cited.