Why the Recycling Industry Must Move Beyond Surface Detection and AI Compensation
The recycling industry is facing a growing challenge. Waste streams are becoming increasingly complex, while at the same time the demand for high-quality recyclates continues to rise. Materials such as black plastics, engineering polymers, multilayer packaging, and additive-rich compounds remain among the most difficult fractions to sort efficiently and economically.
Over the last years, major technology providers including Sesotec, Steinert, Tomra Recycling and others have introduced increasingly sophisticated sorting solutions based on technologies such as Mid-Wave Infrared (MWIR), XRF integration, deep learning, AI-assisted classification systems, and advanced sensor fusion. These developments undoubtedly represent significant engineering progress and have improved sorting performance in many applications.
At the same time, however, they reveal a deeper issue within the industry.
Modern sorting systems are becoming increasingly dependent on software, data processing, and hybrid compensation architectures because the underlying physical sensor information remains fundamentally limited.
The key question is therefore no longer simply whether sorting systems can improve detection rates or sorting purity. The more important question is whether the industry is truly understanding the material itself — or whether increasingly complex software systems are being built to compensate for incomplete sensing.
This problem becomes particularly visible in the case of black plastics. For decades, black plastics have been one of the most difficult material streams to process because traditional NIR systems rely on reflected infrared light to identify polymers. Carbon black pigments absorb most of this light, effectively making the material invisible to conventional optical sorting technologies.
The consequences are well known throughout the recycling industry:
- Low detection rates
- Reduced sorting purity
- Contamination of material streams
- Downgraded recyclates
- Incineration or landfill disposal of otherwise valuable resources
The introduction of MWIR technology was therefore considered a major breakthrough, since carbon black absorption is lower within the MWIR frequency range. Detection improves, certain polymers become easier to distinguish, and recovery rates increase.
Yet even MWIR does not fundamentally eliminate the core limitations that exist under real industrial conditions.
Waste streams in industrial recycling environments are rarely clean, homogeneous, or stable. Plastics are often dirty, aged, mixed, coated, moisture-affected, or filled with additives and reinforcement materials. Under these conditions, signal stability and spectral separability remain difficult despite the advantages of MWIR technology.
Additional challenges persist as well. Polyolefin spectra can still overlap, surface contamination influences sensor readings, and material aging changes spectral behavior over time. The variability of real waste streams reduces reliability and increasingly turns classification into a probabilistic process rather than a deterministic one.
This distinction is critical.
Strong spectral separability under laboratory conditions does not automatically translate into robust industrial sorting performance.
Why AI Is Becoming Central
This is precisely why artificial intelligence is becoming increasingly central within modern recycling systems. Recent developments such as AI-native deep learning platforms and applications like GAINnext™ clearly illustrate the direction the industry is taking.
AI is increasingly used to:
- Interpret weak signals
- Improve classification probabilities
- Compensate for overlapping spectra
- Stabilize sorting decisions
- Adapt systems dynamically to changing waste streams
These approaches undoubtedly improve overall system performance.
However, it is important to understand what this means from a technical perspective.
Artificial intelligence is not removing the sensing limitation itself.
Instead, AI is compensating for incomplete or uncertain sensor information.
This distinction is fundamental because it means the industry is effectively building increasingly sophisticated software layers to overcome the weaknesses of surface-based optical sensing technologies.
At the same time, AI becomes most powerful when combined with richer physical sensing modalities rather than being used primarily to compensate for missing physical information.
AI can amplify physics — but it cannot replace missing physics.
The Rise of Hybrid Compensation Architectures
As a result, modern sorting platforms are evolving into highly complex hybrid architectures that combine MWIR, XRF, AI, deep learning, advanced data fusion, and statistical compensation models.
These systems are technologically impressive, but they also become increasingly complex, expensive, calibration-intensive, software-dependent, and operationally demanding.
AI models often require:
- Continuous retraining
- Large datasets
- Ongoing optimization
- Permanent validation under changing waste conditions
This creates an entirely new operational challenge in which sorting performance becomes heavily dependent on software quality, dataset integrity, and calibration stability.
At the same time, another industrial concern is beginning to emerge:False confidence.
Highly AI-compensated systems may achieve impressive average sorting statistics while still struggling with edge-case materials, unknown formulations, rapidly changing waste streams, or hidden contamination pathways.The challenge is therefore not simply achieving high average sorting performance.
The real challenge is ensuring material reliability under continuously changing real-world conditions.
The Economic Question
The economics of these increasingly hybridized systems also deserve closer attention.
MWIR technology requires:
- Sensitive optics
- Stable operating conditions
- Advanced detectors
- High-performance signal processing
XRF introduces:
- Additional capital costs
- Radiation protection requirements
- Increased maintenance complexity
At the same time, XRF often contributes primarily to bromine detection, heavy-element identification, or flame-retardant recognition rather than fundamentally solving polymer characterization itself.
Meanwhile, AI infrastructure introduces additional software costs, increasing computational requirements, growing data management complexity, and significantly more demanding system integration.
This raises an important industrial question:
At what point does compensation complexity outweigh the economic value created?
Surface Detection Remains the Core Limitation
Despite all technological progress, one fundamental issue remains unresolved:
Most current sorting systems still primarily analyze the surface of the material.
This is becoming increasingly problematic because modern plastics are no longer simple homogeneous polymers. They are highly engineered materials.
Technical plastics frequently contain:
- Talc
- Glass fibers
- Carbon additives
- Mineral fillers
- Reinforcement structures
- Flame retardants
- Multilayer constructions
These internal characteristics ultimately determine:
- Mechanical properties
- Processing behavior
- Recycling suitability
- Material value
- Recyclate quality
Yet surface-based systems often cannot reliably characterize these internal properties.
Two plastics may appear nearly identical externally while having completely different internal compositions and entirely different recycling pathways.
This represents one of the most important unresolved challenges in modern recycling.
Increasingly, recycling performance is determined not by surface chemistry alone, but by volumetric material composition.
The industry is therefore approaching a major technological transition — moving away from simple optical surface recognition and toward deeper material characterization involving:
- Electromagnetic interaction depth
- Dielectric behavior
- Density distribution
- Internal structural heterogeneity
- Volumetric material intelligence
This represents a fundamental conceptual shift:
From surface recognition
toward true material intelligence.
The objective is no longer simply identifying a polymer family.
The objective is understanding the complete material composition and internal structure of the material itself.
Beyond AI Compensation
Artificial intelligence will undoubtedly remain an important component of future recycling systems. But AI alone cannot replace missing physical information.
If the sensor cannot sufficiently access the internal structure of the material, even the most advanced algorithms remain fundamentally limited by the quality and depth of the underlying data.
For this reason, the future of advanced sorting technology may depend less on adding additional compensation layers — and more on fundamentally improving how materials are physically analyzed.
The next major breakthrough may therefore not come from software alone.
It may come from acquiring richer and more meaningful physical information about the material itself.
Conclusion
The latest developments in MWIR, AI-assisted sorting, and hybrid sensor systems undoubtedly represent meaningful progress for the recycling industry. At the same time, however, they also reveal the growing complexity required to compensate for the limitations of surface-based detection technologies.
As plastics become increasingly engineered, reinforced, and compositionally complex, the industry will require technologies capable of:
- Seeing beyond the surface
- Directly understanding internal material composition
- Detecting fillers and multilayer structures
- Characterizing volumetric material behavior
- Delivering this at industrial speed and economically viable scale
The future of recycling will therefore not be defined solely by better algorithms.
It will be defined by better material understanding.
Because ultimately, successful circularity depends not just on detecting plastics — but on truly understanding what they are made of.
Author: Eric van Looy