Seeing the Invisible: The Physics, Economics, and Intelligence of Black Plastic Recovery

A Whitepaper by Veriprajna | Deep AI Solutions

Executive Summary

The global circular economy faces a systemic failure that is invisible to the human eye but catastrophic to the environment and the balance sheet alike. It is a failure that occurs millions of times daily within the intricate machinery of Materials Recovery Facilities (MRFs) worldwide. When a conscientious consumer washes a black plastic food tray, sorts a dark coffee lid, or discards a black automotive component, they operate under the assumption that the recycling infrastructure is capable of recovering these materials. This assumption is technologically unfounded for a staggering volume of waste.

Millions of tons of black plastics—primarily Polypropylene (PP), Polyethylene (PE), Polystyrene (PS), and Acrylonitrile Butadiene Styrene (ABS)—are systematically ejected from the recycling stream and sent to landfill or incineration every year. ¹ This is not due to a lack of material value; recycled black pellets command market prices rivaling, and occasionally exceeding, virgin resins. ³ Nor is it due to the chemical impossibility of recycling these polymers. The failure is strictly one of sensor physics .

The modern recycling industry relies on optical sorting technology underpinned by Near-Infrared (NIR) spectroscopy. This technology identifies materials by analyzing light reflected off waste objects. However, the carbon black pigment used to color approximately 15% of the waste stream is a potent broadband absorber of light. ¹ To a standard NIR sensor, a black plastic tray on a black conveyor belt generates zero contrast. It creates a data void—a blind spot where the sorting algorithms perceive nothing but background noise. Consequently, the pneumatic ejectors remain silent, and the valuable polymer falls into the residue stream, representing a massive leakage of economic value and a profound environmental liability.

This whitepaper posits that this crisis cannot be resolved through the current trend of "AI Wrappers"—thin software layers built atop generalized Large Language Models (LLMs). One cannot prompt a generative model to hallucinate physical data that was never captured by the sensor. If the contrast is absent, the AI is blind. The solution requires a fundamental shift in the data acquisition layer, moving from the visible and near-infrared spectrum into the Mid-Wave Infrared (MWIR) thermal domain, coupled with a shift in machine intelligence from 2D spatial recognition to 1D-Convolutional Neural Networks (1D-CNNs) designed for spectral signal processing.

Veriprajna positions itself as a Deep Tech solution provider, bridging the chasm between advanced sensor physics and industrial artificial intelligence. We do not simply process images; we engineer the capture of chemical reality. By implementing Hyperspectral Imaging (HSI) in the 2.7–5.3 µm range ⁶, we render carbon black transparent, revealing the unique chemical fingerprints of the polymers beneath. This document outlines the comprehensive engineering architecture, the economic imperatives, and the strategic vision required to close the loop on the darkest fraction of the global waste stream.

Part I: The Black Hole in the Circular Economy

To understand the magnitude of the technological challenge Veriprajna addresses, one must first quantify the scale of the failure. The global production of plastics has surged relentlessly, doubling in the last two decades to approximately 460 million tonnes annually. ⁸ Within this torrent of material, black plastics represent a significant, albeit often unquantified, "dark matter" of the waste stream.

1.1 The Ubiquity and Utility of Black Polymers

Black plastics are not an accidental byproduct; they are a deliberate engineering choice deeply embedded in the manufacturing supply chain. In the automotive sector, black Polypropylene (PP) and ABS are the standards for interior dashboards, bumpers, and structural components due to their UV stability and ability to mask dirt and scratches. In the electronics industry, black casings for computers, televisions, and peripherals are ubiquitous.

Crucially, in the fast-moving consumer goods (FMCG) and food packaging sectors, black plastics—often CPET (Crystalline Polyethylene Terephthalate) or PP—are prized for their aesthetic neutrality. They provide a high-contrast background that makes food colors appear more vibrant on supermarket shelves. ¹⁰ Furthermore, the color black allows manufacturers to utilize "jazz" or mixed-color recycled feedstock. By adding carbon black pigment, producers can homogenize a heterogeneous mix of multi-colored recycled plastics into a uniform, sleek product, effectively masking imperfections. ⁶

However, this utility creates a paradoxical dead-end. The very pigment that allows for the usage of recycled content—Carbon Black—renders the final product undetectable to standard recycling machinery. This results in a "single-use" trap for materials that are chemically capable of infinite circularity.

1.2 The Statistical Reality of Loss

The global statistics on plastic waste management paint a grim picture of inefficiency. Of the

353 million tonnes of plastic waste generated annually, only 9% is successfully recycled. ⁸ The fate of the remainder is a mix of landfilling (50%), incineration (19%), and uncontrolled mismanagement (22%). ⁸

Black plastics disproportionately populate the non-recycled fractions. Estimates suggest that black packaging constitutes between 3% and 15% of the total plastic waste stream depending on the region. ⁵ In a typical Materials Recovery Facility (MRF) processing 50,000 tons annually, this translates to thousands of tons of lost material. This is not merely trash; it is lost revenue. With recycled pellet prices for polymers like ABS and PP hovering between $1,000 and $1,200 per ton ⁴, the inability to sort black plastic represents a multi-million dollar revenue leak for every major operator.

Moreover, the environmental cost is compounded by the "downcycling" phenomenon. Because black plastics cannot be sorted by polymer type (e.g., separating PP from PS), even when they are manually recovered, they are often aggregated into a low-value mixed rigid stream. This mixed stream has limited utility and is often destined for low-grade applications like park benches or, worse, incineration for energy recovery, which releases the embodied carbon of the fossil-fuel-based polymers back into the atmosphere. ¹

1.3 The Regulatory Siege

The "laissez-faire" approach to this waste stream is ending. Regulatory frameworks across the globe are tightening the noose on unrecyclable packaging. The European Union’s Packaging and Packaging Waste Regulation (PPWR) is perhaps the most aggressive, mandating that all packaging on the EU market must be recyclable by 2030. ¹³

This regulation moves the definition of "recyclable" from a theoretical possibility to a practical reality. It is no longer sufficient to say that Black PP can be recycled in a lab; it must be proven that it is recycled at scale in industrial facilities. If the standard optical sorters cannot see it, it is legally classified as unrecyclable. This poses an existential threat to brands using black packaging, forcing them to either abandon the color (incurring rebranding costs) or demand a technological solution that makes their packaging visible.

Simultaneously, the economic penalties for failure are rising. Landfill taxes in Europe have soared, with costs in some member states exceeding €100 per ton to discourage disposal. ¹⁵ In the United States, while landfilling remains cheaper, tipping fees are rising, and Extended Producer Responsibility (EPR) laws in states like California (SB 54) are shifting the financial burden of waste management onto the producers. ¹⁷ The "free" disposal of black plastic is a relic of the past.

Part II: The Physics of Blindness

To solve the black plastic problem, one must move beyond the superficial understanding of "trash" and engage with the fundamental physics of light-matter interaction. The failure of the current recycling infrastructure is not a failure of robotics or mechanics; it is a failure of spectroscopy.

2.1 The Carbon Black Pigment

The villain in this narrative, if there is one, is Carbon Black. It is a material produced by the incomplete combustion of heavy petroleum products. Its defining characteristic is its extremely high absorption coefficient across a vast swath of the electromagnetic spectrum. It absorbs ultraviolet light, visible light (hence its black appearance), and crucially, Near-Infrared (NIR) light. ¹

This is distinct from other colorants. Organic dyes, for instance, might appear black to the human eye because they absorb red, green, and blue light, but they are often transparent in the infrared spectrum. Carbon black, however, acts as a photon trap. When light photons strike a carbon-black pigmented surface, their energy is dissipated as heat rather than reflected back to the source.

2.2 The Limitations of Standard NIR Spectroscopy

The recycling industry’s workhorse sensor is the NIR spectrometer, typically operating in the range of 900 nm to 1700 nm (0.9–1.7 µm). ¹⁹

●​ The Mechanism: In a standard sorting unit (e.g., from manufacturers like Tomra or Steinert), halogen lamps illuminate the conveyor belt. The light strikes the plastic objects.

●​ The Signal: For a clear or colored plastic (like a blue detergent bottle), the light penetrates the surface, interacts with the molecular bonds (C-H, N-H, O-H), and is reflected back with specific wavelengths absorbed. This absorption pattern is the "spectral fingerprint."

●​ The Failure Mode: When the light strikes a black plastic tray, the carbon black pigment absorbs the incident radiation before it can interact significantly with the polymer matrix. The sensor receives effectively zero return signal.

This phenomenon creates a fundamental contrast failure. The conveyor belts used in these facilities are typically black rubber. A black object on a black belt, under a sensor that sees no reflection from either, results in a "flatline" signal. The sorting computer cannot distinguish the object from the background. To the robot, the belt appears empty. ²

2.3 The "Wrapper" Trap: Why LLMs Cannot Fix Physics

We currently exist in a technological epoch defined by the ascendancy of Large Language Models (LLMs) and Generative AI. A proliferation of "AI Consultancies" has emerged, offering to solve enterprise problems by "wrapping" models like GPT-4 or Claude via API. ²¹ These "Wrapper" businesses operate on the premise that sufficient prompting and context can extract intelligence from any data stream.

However, applying this logic to the physical sorting of black plastics is a category error. Generative AI models are probabilistic engines designed to predict the next token in a sequence or the next pixel in an image based on vast training datasets. They require input data to function.

●​ The Input Void: In the case of black plastic sorting via standard NIR, the input data is a null set. There is no spectral curve to analyze; there is only noise.

●​ Hallucination Risk: If one were to force an LLM to classify this noise, it might hallucinate a result based on priors (e.g., "most trays are PP"), but this is not sensing; it is guessing. In an industrial process requiring 99% purity to prevent contamination, guessing is unacceptable.

●​ Latency Unsuitability: Industrial sorting requires decisions in the millisecond range (<10-20ms). Sending data to a cloud-based LLM API introduces latency in the range of hundreds of milliseconds to seconds, rendering real-time actuation impossible. ²³

Veriprajna asserts that the solution lies in Deep Tech —the rigorous application of scientific discovery to engineering problems—rather than the convenient application of pre-trained models. We do not try to make the AI "smarter" at guessing; we change the physics of the sensor to make the invisible visible.

Part III: The Photonic Solution – Mid-Wave Infrared (MWIR)

If the data does not exist in the Near-Infrared, we must shift our gaze to a different window of the electromagnetic spectrum. The solution lies in the Mid-Wave Infrared (MWIR) region, specifically between 2.7 µm and 5.3 µm.

3.1 The Electromagnetic Window

Carbon black’s absorption is not infinite. As the wavelength of light increases, the absorption coefficient of carbon black decreases. By moving from the Short-Wave Infrared (SWIR) into the Mid-Wave Infrared (MWIR), we cross a threshold where the pigment becomes sufficiently transparent to allow for spectral analysis. ⁶

●​ NIR/SWIR (0.9 – 2.5 µm): Dominated by "overtones" of molecular vibrations. These are weak signals, easily swamped by carbon black absorption.

●​ MWIR (2.7 – 5.3 µm): This region contains the fundamental vibrations of the polymer molecules. The C-H stretching vibrations, C=O carbonyl stretches, and other bond energies resonate strongly in this frequency.

In the MWIR domain, the spectral signatures of plastics are orders of magnitude stronger than in the NIR. This increased signal strength, combined with the reduced absorption of the pigment, allows the sensor to "look through" the darkness. A black Polypropylene tray, which is a void in the NIR, glows with a distinct, jagged spectral line in the MWIR, clearly distinguishable from a black Polyethylene or Polystyrene object. ²⁵

3.2 Hardware Architecture: The Specim FX50

Veriprajna builds its solutions upon best-in-class hardware, specifically the Specim FX50 hyperspectral camera, which is currently the only commercially viable instrument covering the full 2.7–5.3 µm range necessary for this application. ⁷

The engineering complexity of this sensor highlights the "Deep Tech" nature of the solution:

●​ Detector Material: Unlike the Silicon sensors used in RGB cameras (cheap, uncooled) or the InGaAs sensors used in NIR (moderately expensive), MWIR requires exotic semiconductor materials like Indium Antimonide (InSb) or Mercury Cadmium Telluride (MCT) . ⁷ These materials are sensitive to thermal radiation.

●​ Cryogenic Cooling: MWIR sensors are essentially detecting heat. To prevent the sensor from being blinded by its own internal heat generation, the detector must be cooled to cryogenic temperatures (approx. 77 Kelvin). The FX50 integrates a Stirling cooler—a mechanical refrigeration unit—directly into the camera body. This cooling requirement significantly increases the barrier to entry, distinguishing serious industrial players from hobbyist attempts. ²⁷

●​ Spectral Resolution: The camera acts as a line-scanner (push-broom), capturing 154 distinct spectral bands for every pixel in the line. With a spectral sampling of ~8.44 nm and a spatial resolution of 640 pixels, it generates a massive, high-dimensional data cube ($x, y, \lambda$) in real-time. ⁷

3.3 Seeing "Chemistry" Not "Color"

The shift to Hyperspectral Imaging (HSI) represents a paradigm shift from "Computer Vision" to "Chemical Vision."

●​ Computer Vision (RGB): Analyzes the external appearance—shape, texture, label, color.

●​ Hyperspectral Imaging (HSI): Analyzes the intrinsic composition. When the FX50 scans a mixed stream of black plastics, it does not see "black shapes." It sees a stream of chemical data.

●​ Polyethylene (PE): Characterized by strong C-H stretching absorptions near 3.4 µm.

●​ Polystyrene (PS): Shows distinct aromatic C-H stretching modes that are sharp and unique, clearly separable from the aliphatic C-H bands of PE/PP. ⁶

●​ Flame Retardants: Crucially, MWIR can often detect the presence of flame retardants in ABS (used in electronics), which is vital for separating WEEE (Waste Electrical and Electronic Equipment) plastics from standard packaging plastics. ²⁵

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

Part IV: The Cognitive Engine – 1D-Convolutional Neural Networks

Capturing the MWIR Hypercube is only half the battle. The sensor generates gigabytes of data per second. To make this actionable, we need an Artificial Intelligence architecture capable of interpreting these complex chemical signatures with near-zero latency. This is where Veriprajna’s proprietary 1D-Convolutional Neural Networks (1D-CNNs) come into play.

4.1 The Failings of 2D Computer Vision in Waste

The standard approach in AI image recognition is the 2D-CNN (e.g., ResNet, YOLO). These networks are designed to recognize spatial features: the curve of a cat's ear, the rectangle of a license plate.

●​ Spatial Irrelevance: In waste recycling, spatial features are unreliable. A plastic bottle might be crushed flat, twisted, torn, or dirty. A fragment of a black car bumper looks spatially identical to a fragment of a black food tray. Relying on "shape" leads to poor accuracy. ²

●​ Spectral Blindness: Applying 2D-CNNs to hyperspectral data often treats the spectral bands as mere "color channels" (like R, G, B), failing to capture the subtle, high-frequency relationships between the 154 spectral bands that define the chemical fingerprint.

4.2 The 1D-CNN Paradigm

Veriprajna treats the problem not as image recognition, but as signal processing . For every pixel on the conveyor belt, we extract a 1D vector of 154 values—the spectrum. We feed this vector into a 1D-CNN. ³¹

Architectural Specifics:

1.​ Input Layer: A $1 \times 154$ vector representing the normalized reflectance/radiance values.

2.​ Convolutional Layers (Conv1D): Instead of square kernels sliding over an image, we use linear kernels sliding over the spectrum. These kernels learn to detect specific spectral shapes: a sharp drop at 3.4 µm, a broad shoulder at 4.0 µm, or a specific doublet peak. The network learns the "grammar" of molecular bonds. ³³

3.​ Pooling Layers: Max-pooling reduces the dimensionality, making the model robust to minor shifts in calibration or noise. ³⁵

4.​ Fully Connected Layers: These aggregate the features extracted by the convolutional layers to map them to specific classes (e.g., Black PP, Black PE, Black ABS).

5.​ Output: A softmax probability distribution (e.g., [PP: 99%, PE: 0.5%, Other: 0.5%]).

4.3 Superiority Over Transformers and LSTMs

In the academic literature, there is significant interest in using Transformers (Self-Attention mechanisms) and LSTMs (Long Short-Term Memory networks) for spectral analysis. ³² While these are powerful, they are often ill-suited for the harsh reality of industrial sorting.

●​ Latency: Transformers have a quadratic computational complexity $O(N^2)$ with respect to sequence length. LSTMs are sequential and difficult to parallelize. Both introduce inference latencies that can exceed the strict time budget of a high-speed sorting line (belt speeds of 3m/s). ³²

●​ 1D-CNN Efficiency: 1D-CNNs have a linear complexity $O(N)$ and are highly parallelizable on GPUs. They can be optimized using TensorRT to run at thousands of frames per second on edge hardware, ensuring that the classification keeps pace with the physical flow of waste. ³⁶

Veriprajna’s models are not just accurate; they are fast . In the time it takes a Transformer to "attend" to the global context of a spectrum, our 1D-CNN has already classified the pixel and triggered the air jet.

Part V: System Architecture & Integration

Veriprajna provides the "brain," but this brain must reside within a body. The integration of MWIR HSI into a robotic sorting line requires a sophisticated sensor fusion and edge computing architecture.

5.1 Sensor Fusion: The Best of Both Worlds

While MWIR is superior for material classification, it has lower spatial resolution than standard RGB cameras and is expensive. To optimize performance and cost, we employ Sensor Fusion . ³⁸

●​ The RGB Stream: A high-resolution, low-cost RGB camera scans the belt. Its job is solely Segmentation —detecting the boundaries of objects (e.g., "There is an item located at pixels 100-200"). It creates a "mask" of where the waste is.

●​ The MWIR Stream: The Specim FX50 captures the spectral data.

●​ The Fusion Engine: The system overlays the RGB mask onto the MWIR data. It queries the spectral data inside the object boundaries identified by the RGB camera. The 1D-CNN classifies the material based on the average or pixel-wise spectrum within that mask.

●​ The Result: The robot receives a composite data packet: "Object ID #452 is a Black Polypropylene Tray, located at $(x,y)$, orientation $\theta$."

This hybrid approach allows us to use the high spatial fidelity of RGB to guide the precise grasping or ejection, while relying on the high spectral fidelity of MWIR for the decision-making. ⁴⁰

5.2 Edge Computing and Real-Time Actuation

There is no cloud in this loop. The bandwidth required to stream raw hyperspectral video (hundreds of bands at high frame rates) is massive, and the latency of internet round-trips is unacceptable.

●​ Edge AI: Veriprajna deploys industrial-grade Edge GPUs (e.g., NVIDIA Jetson AGX Orin or robust IPCs with RTX A-series cards) directly on the sorting machine.

●​ Inference Speed: Our optimized 1D-CNN models run with inference times below 5 milliseconds.

●​ Actuation: The system interfaces directly with the Programmable Logic Controller (PLC) of the sorter.

○​ Optical Sorters: For light packaging (films, flakes), the system triggers a precise blast of compressed air from a valve bar to eject the target material. ¹¹

○​ Robotic Arms: For rigid plastics or heavier items, the system guides high-speed Delta robots ("Spider robots") to pick the black plastic item and place it in the correct bin. ⁴³

5.3 The Ecosystem of Machinery

Veriprajna works within the existing ecosystem of sorting hardware. Our technology is designed to integrate with or upgrade systems from major manufacturers like Tomra (e.g., Autosort Black) and Steinert (e.g., Unisort BlackEye). ⁴² While these manufacturers offer their own proprietary solutions, Veriprajna offers a specialized, customizable software layer that can adapt to specific, difficult waste streams (e.g., specific flame-retardant mixes or novel multilayer films) that standard firmware may miss. We provide the "bespoke intelligence" for the standardized hardware.

Part VI: The Economic Imperative

The adoption of Veriprajna’s technology is driven not just by environmental conscience, but by hard economic metrics. The recovery of black plastic transforms a cost center into a profit center.

6.1 The "Green Premium" and Commodity Prices

Recycled plastic pellets (Post-Consumer Recycled or PCR) have decoupled from the price of virgin oil-based resins. Driven by corporate sustainability goals and regulatory mandates, high-quality PCR often trades at a premium.

●​ Black Recycled Pellets:

○​ Recycled PP (rPP): Trades in the range of $1,130 – $1,200 per ton . ⁴

○​ Recycled ABS (rABS): Trades between $800 and $1,100 per ton . ³

○​ Virgin vs. Recycled: In Europe, recycled plastics have reached price parity with virgin plastics, and in some high-demand sectors, they command a premium due to scarcity. ⁴⁶

A facility that discards 10% of its input as "residue" because it is black is essentially throwing away money. For a plant processing 50,000 tons a year, recovering just 2,000 tons of Black PP represents over $2 million in potential revenue .

6.2 The Cost of Waste: Landfill and Carbon Taxes

The alternative to recovery—disposal—is becoming prohibitively expensive.

●​ Landfill Tipping Fees: In the US, the average tipping fee is ~$60/ton, but this masks huge regional variances (>$100/ton in the Northeast and West Coast). ⁴⁷ In the EU, landfill bans and taxes push these costs even higher to force circularity. ¹⁶

●​ Incineration and ETS: As the EU Emissions Trading System (ETS) expands to cover waste incineration, facilities will be charged for the CO2 emitted when burning plastics. Since plastic is essentially solid fossil fuel, the carbon cost is significant. Recovering the plastic avoids this tax. ¹⁴

6.3 ROI Calculation

Case Study: Mid-Sized MRF (Europe)

●​ Throughput: 50,000 tons/year.

●​ Black Plastic Content: 5% (2,500 tons).

●​ Current Fate: Incineration (Cost: €100/ton gate fee + tax) = -€250,000/year cost.

●​ With Veriprajna + MWIR System:

○​ Recovery Rate: 90% (2,250 tons).

○​ Revenue: Sold at €900/ton = +€2,025,000 revenue.

○​ Disposal Savings: €225,000 avoided cost.

○​ Total Annual Impact: €2.25 Million .

●​ System CAPEX: ~$300,000 (Hardware + Integration).

●​ Payback Period: < 2 Months .

The economics are undeniable. Deep Tech pays for itself.

Part VII: Strategic Positioning – The "Deep Tech" Moat

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

7.1 Escaping the "Wrapper" Trap

The "Wrapper" business model—building a UI on top of someone else's API—is structurally fragile. It has no "moat." If the underlying model provider (e.g., OpenAI) updates their product to include your feature, your business evaporates. Furthermore, wrappers are constrained by the limitations of the underlying model. You cannot "prompt" GPT-4 to see MWIR spectra it was never trained on. ²¹

Veriprajna is a Deep Tech Company.

●​ Proprietary Data: We generate our own data. The MWIR spectral signatures of dirty, crushed, real-world waste are not available on the public internet. We build and own this dataset.

●​ Physical Moat: Our solution requires the integration of cryogenic hardware, edge computing, and robotics. This cannot be copied by a couple of developers in a coffee shop. It requires engineering rigor and physical access to industrial sites. ⁵⁰

7.2 Vertical Integration and Partnership

We do not operate in a vacuum. We position ourselves as the Intelligence Layer within the hardware ecosystem.

●​ Sensor Agnostic but Opinionated: While we prefer the Specim FX50 for its performance, our 1D-CNN architectures can be adapted to other MWIR sensors as they emerge.

●​ Retrofit Capability: A key part of our strategy is retrofitting existing optical sorters.

Many MRFs have perfectly good conveyor systems and air ejectors but outdated sensor heads. Veriprajna can upgrade the "eyes and brain" of the machine without replacing the "body," lowering the CAPEX for our clients.

7.3 The Future: Beyond Sorting

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

●​ Digital Product Passports (DPP): We can track exactly what materials are passing through the recycling chain.

●​ Quality Assurance: We can certify the purity of a bale of recycled plastic before it is shipped to a converter, reducing rejected loads.

●​ Brand Audits: We can provide data to producers on how much of their specific black packaging is actually being recovered, helping them meet EPR reporting requirements.

Conclusion

"You wash your recycling. You sort it. And the AI throws it in the trash anyway."

This statement encapsulates the tragedy of the current system. It is a failure of vision—both literal and metaphorical. By relying on the visible spectrum, we have blinded ourselves to the value hidden in the dark.

The solution is not more of the same. It is not a better prompt, a larger language model, or a slicker user interface. The solution is physics . It is the rigorous application of Mid-Wave Infrared Hyperspectral Imaging to render the invisible visible. It is the application of specialized 1D-Convolutional Neural Networks to interpret the chemical reality of our waste.

Veriprajna stands at this intersection of photonics and intelligence. We do not just process data; we engineer the acquisition of truth. By enabling the recovery of black plastics, we not only close a massive loop in the circular economy but also demonstrate the power of Deep Tech to solve the hard, physical problems of our time.

We invite you to stop looking at the picture, and start looking at the chemistry.

#GreenTech #Recycling #ComputerVision #Sustainability #DeepTech

Appendix: Technical Specifications

Table 1: Veriprajna MWIR System Specifications

Component	Specifci ation	Rationale
Sensor Core	Specim FX50 (InSb/MCT)	Only commercial sensor covering full 2.7–5.3 µm range.7
Spectral Bands	154 Bands	High resolution for diferentiating chemically similar polymers (e.g., PE vs PP).

Cooling	Stirling Cooler (Cryogenic)	Essential to minimize thermal noise in MWIR range.27
Frame Rate	380 Hz (Full Frame)	Supports conveyor speeds >2 m/s with high overlap.
AI Model	Custom 1D-CNN (5-Layer)	Optimized for <5ms latency on Edge GPU; high spectral feature extraction.
Inference Hardware	NVIDIA Jetson AGX Orin / RTX 4090	Edge computing required to eliminate network latency.23
Detectable Materials	Black PP, PE, PS, ABS, PVC, Rubber	Covers >90% of black plastic waste stream.
Connectivity	GigE Vision / GenICam	Standard industrial interface for seamless integration.

Works cited

1. Transforming Black Plastic Recycling, accessed December 12, 2025, https://www.plasticsengineering.org/2025/07/transforming-black-plastic-recycling-005543/

2. Seeing The Unseen: Black Plastics Detection & Machine Vision, accessed December 12, 2025, https://www.portableas.com/news/black-plastics-detection-machine-vision/

3. recycled abs plastic material price, accessed December 12, 2025, https://www.ecoplashk.com/productstags/recycled-abs-plastic-material-price.html

4. Polypropylene Price per kg in 2025: Trends, Forecasts, and Market Analysis - Sales Plastics, accessed December 12, 2025, https://salesplastics.com/polypropylene-price-per-kg/

5. Black plastics recycling: towards a circular economy, accessed December 12, 2025, https://www.recycling-magazine.com/2022/09/22/black-plastics-recycling-towards-a-circular-economy/

6. Overcoming Black Plastic Recycling Challenge with MWIR Hyperspectral Imaging Technology - Qd-uki.co.uk, accessed December 12, 2025, https://qd-uki.co.uk/overcoming-black-plastic-recycling-challenge-with-mwir-hyperspectral-imaging-technology/

7. Specim FX50 - Hyperspectral Camera - MEETOPTICS, accessed December 12, 2025, https://www.meetoptics.com/detection-devices/cameras/hyperspectral-camera/s/specim/p/Specim%20FX50

8. Plastic pollution is growing relentlessly as waste management and recycling fall short, says OECD, accessed December 12, 2025, https://www.oecd.org/en/about/news/press-releases/2022/02/plastic-pollution-is-growing-relentlessly-as-waste-management-and-recycling-fall-short.html

9. Plastic Pollution - Our World in Data, accessed December 12, 2025, https://ourworldindata.org/plastic-pollution

10. Why is black plastic packaging so hard to recycle? - The World Economic Forum, accessed December 12, 2025, https://www.weforum.org/stories/2019/12/black-plastic-recycling-supermarkets-waste/

11. UniSort Black for reliable detection and sorting of different materials - steinert, accessed December 12, 2025, https://steinertglobal.com/us/sorting-systems/sensor-sorting/nir-sorting-systems/unisort-black/

12. High Quality Recycled ABS Granules at A Competitive Price - Exir Polymer, accessed December 12, 2025, https://exirpolymer.com/granules/abs-granules/

13. Plastic Recycling with NIR Technology: A Complete Guide - Solid Scanner, accessed December 12, 2025, https://solidscanner.com/basics-and-limitations-of-nirs-and-how-it-is-used-in-plastic-recycling/

14. Europe's Drive to Slash Plastic Waste Moves Into High Gear - Yale E360, accessed December 12, 2025, https://e360.yale.edu/features/europes-drive-to-slash-plastic-waste-moves-into-high-gear

15. Final layout - MRBT costs study Apr23 - Zero Waste Europe, accessed December 12, 2025, https://zerowasteeurope.eu/wp-content/uploads/2023/04/ZWE_Apr23_Study_MRBT_costs_study.pdf.pdf

16. Economic instruments and separate collection systems — key strategies to increase recycling | Publications - European Environment Agency, accessed December 12, 2025, https://www.eea.europa.eu/publications/economic-instruments-and-separate-collection

17. Can recycling advancements get black rigid plastics off packaging's blacklist?, accessed December 12, 2025, https://www.packagingdive.com/news/recycling-technology-advancements-black-rigid-plastic-packaging/759439/

18. Why Is Carbon Black a Particular Challenge for Automated Plastic Sorting Technologies?, accessed December 12, 2025, https://lifestyle.sustainability-directory.com/learn/why-is-carbon-black-a-particular-challenge-for-automated-plastic-sorting-technologies/

19. Real-Time, Non-Contact Sorting of Plastics for Recycling Using Hyperspectral Imaging - Headwall Photonics, accessed December 12, 2025, https://headwallphotonics.com/application-notes/real-time-non-contact-sorting-of-plastics-for-recycling-using-hyperspectral-imaging/

20. What Are the Limitations of Current Near-Infrared (NIR) Sorting Technology for Mixed Plastics? → Learn - Lifestyle → Sustainability Directory, accessed December 12, 2025, https://lifestyle.sustainability-directory.com/learn/what-are-the-limitations-of-current-near-infrared-nir-sorting-technology-for-mixed-plastics/

21. Beyond the Blank Slate: Escaping the AI Wrapper Trap - jeffreybowdoin.com, accessed December 12, 2025, https://jeffreybowdoin.com/beyond-blank-slate-escaping-ai-wrapper-trap/

22. AI Wrappers - The Quiet Race for Interface Dominance - The Prompt Engineering Institute, accessed December 12, 2025, https://promptengineering.org/ai-wrappers-the-quiet-race-for-interface-dominance-2/

23. The AI infrastructure reckoning: Optimizing compute strategy in the age of inference economics - Deloitte, accessed December 12, 2025, https://www.deloite.com/us/en/insights/topics/technology-management/tech-tretnds/2026/ai-infrastructure-compute-strategy.html

24. Black plastic identification by hyperspectral imaging in mid-wave infrared PubMed, accessed December 12, 2025, https://pubmed.ncbi.nlm.nih.gov/41067168/

25. Revolutionizing Black Plastic Sorting: The Power of the Specim FX50 MWIR Hyperspectral Camera and Its Integration into Industria - ATEsystem, accessed December 12, 2025, https://img.eshop.atesystem.cz/fotocache/files/Dokumenty/Revolutionizing-Black-Plastic-Sorting.pdf

26. Hyperspectral Imaging: Shaping the Future of Plastic Recycling - Konica Minolta Sensing, accessed December 12, 2025, https://sensing.konicaminolta.us/us/blog/hyperspectral-imaging-shaping-the-future-of-plastic-recycling/

27. Specim FX50 (2.7-5.3 µm) - Channel Systems, accessed December 12, 2025, https://channelsystems.ca/products/midwave-infrared-mwir/specim-fx50-27-53-um

28. Specim FX50 hyperspectral camera - A breakthrough in black plastic sorting, accessed December 12, 2025, https://www.specim.com/breakthrough-in-black-plastic-sorting/

29. SPECIM FX50 - デルフトハイテック株式会社, accessed December 12, 2025, https://dht.co.jp/wp/wp-content/uploads/2019/06/Specim-FX50-Technical-Datasheet-01_dht.pdf

30. The Differences Between SWIR, MWIR, and LWIR Cameras - Axiom Optics, accessed December 12, 2025, https://www.axiomoptics.com/blog/the-diferences-between-swir-mwir-and-lwirf-cameras/

31. Deep&Dense Convolutional Neural Network for Hyperspectral Image Classification - MDPI, accessed December 12, 2025, https://www.mdpi.com/2072-4292/10/9/1454

32. (PDF) Hyperspectral Image Classification: An Analysis Employing CNN, LSTM, Transformer, and Attention Mechanism - ResearchGate, accessed December 12, 2025, https://www.researchgate.net/publication/369128397_Hyperspectral_Image_Classification_An_Analysis_Employing_CNN_LSTM_Transformer_and_Atention_Mechatnism

33. Hands-On Fundamentals of 1D Convolutional Neural Networks—A Tutorial for Beginner Users - MDPI, accessed December 12, 2025, https://www.mdpi.com/2076-3417/14/18/8500

34. 1D-CNN architecture (based in ¹⁴). We named this... - ResearchGate, accessed December 12, 2025, https://www.researchgate.net/figure/D-CNN-architecture-based-in-14-We-named-this-architecture-1D-CNN-O-O-stands-for_fig2_369128397

35. 5. Convolutional 1D Network Classification - ReCoDE- AI For Patents, accessed December 12, 2025, https://imperialcollegelondon.github.io/ReCoDE-AIForPatents/5_Convolutional_1D_Network_Classification/

36. Hyperspectral Image Classification: An Analysis Employing CNN, LSTM, Transformer, and Attention Mechanism - IEEE Xplore, accessed December 12, 2025, https://ieeexplore.ieee.org/document/10064297/

37. [1903.09834] 1D-Convolutional Capsule Network for Hyperspectral Image Classification, accessed December 12, 2025, https://arxiv.org/abs/1903.09834

38. Enhanced Waste Sorting Technology by Integrating Hyperspectral and RGB Imaging - MDPI, accessed December 12, 2025, https://www.mdpi.com/2313-4321/10/5/179

39. (PDF) Enhanced Waste Sorting Technology by Integrating Hyperspectral and RGB Imaging, accessed December 12, 2025, https://www.researchgate.net/publication/395723120_Enhanced_Waste_Sorting_Technology_by_Integrating_Hyperspectral_and_RGB_Imaging

40. FusionSort: Enhanced Cluttered Waste Segmentation with Advanced Decoding and Comprehensive Modality Optimization The authors would like to thank MBZUAI for providing the computational resources used in this research. We also gratefully acknowledge the Multispectral Waste (MSWaste) dataset team for sharing their dataset. - arXiv, accessed December 12, 2025, https://arxiv.org/html/2508.19798v1

41. Specim Case Study: Waste Robotics - Hyperspectral Imaging: A Tool for Future Waste Management - YouTube, accessed December 12, 2025, https://www.youtube.com/watch?v=9gqLT70auzw

42. AUTOSORT™ BLACK: Advanced Sorting for Black Plastics - TOMRA, accessed December 12, 2025, https://www.tomra.com/waste-metal-recycling/products/machines/autosort-black

43. Revolutionising the sorting of black plastics in industry with the Specim FX50 MWIR hyperspectral camera - OptiXs, accessed December 12, 2025, https://www.optixs.cz/en/revolucni-trideni-cernych-plastu-v-prumyslu-s-hyperspektralni-kamerou-specim-fx50-mwir

44. STEINERT UniSort sorting technology for black plastics - recovery Magazine, accessed December 12, 2025, https://www.recovery-worldwide.com/en/artikel/steinert-unisort-sorting-technology-for-black-plastics-3932622.html

45. Pp Pellets Price - Made-in-China.com, accessed December 12, 2025, https://www.made-in-china.com/products-search/hot-china-products/Pp_Pellets_Price.html

46. Impact on Virgin vs. Recycled Plastics Prices and Implications for a Production Cap - IEEFA, accessed December 12, 2025, https://ieefa.org/sites/default/files/2024-11/Reviewed-14920-Briefing%20note_Petchem%20recycling%20prices.pdf

47. US landfilling costs jumped sharply last year - Resource Recycling, accessed December 12, 2025, https://resource-recycling.com/recycling/2023/06/12/us-landfilling-costs-jumped-sharply-last-year/

48. The Deep Tech Opportunity - Matthew Mandel, accessed December 12, 2025, https://matthewmandel.com/2025/01/08/the-deep-tech-opportunity/

49. Myth of "Unfundable" LLM Wrappers – Build Profitable AI SaaS | 1Mby1M | 1m1m, accessed December 12, 2025, https://1m1m.sramanamitra.com/virtual-accelerator/no-equity/the-myth-of-unfundable-llm-wrapper-startups/

50. Wrappers, deeptechs, and generative AI: a profitable but fragile house of cards, accessed December 12, 2025, https://www.duperrin.com/english/2025/05/20/wrappers-deeptechs-generative-ai/