AI system for the circular economy
The linear "extract-produce-dispose" model creates $1,000 worth of waste annually in the EU manufacturing sector alone. The transition to a circular economy is an engineering challenge: how to track materials throughout their entire lifecycle, predict when they'll be returned, and optimize remanufacturing. AI solves tracking, forecasting, and optimization problems that are impossible with a manual approach.
Tracking of materials and products
Digital Product Passport (DPP)
Starting in 2026, the Ecodesign Regulation (ESPR) will require manufacturers of electronics, batteries, and textiles to issue a Digital Product Passport—a machine-readable profile with material composition, emissions, and disassembly instructions. AI component: automatic generation of DPPs from BOM (Bill of Materials) and lifecycle assessment data.
The LLM agent bypasses ERP (SAP MM, Oracle) and Product Lifecycle Management (Siemens Teamcenter, PTC Windchill), extracts the material composition, calculates LCA using the Ecoinvent database, and generates DPP in GS1 Digital Link + JSON-LD format. BOM coverage → DPP without manual entry: 78% of components in the e-company pilot.
Reverse logistics optimization
Product lifecycle end: when, where, and how many units will be returned? Time-series forecasting (TFT — Temporal Fusion Transformer) based on historical returns data, taking into account the sale date, region, product type, and economic indicators. A MAPE of 14% over a six-month horizon is sufficient for reprocessing facility capacity planning.
Optimization remanufacturing
Sorting and condition assessment
Returned products need to be quickly classified: reuse as is / refurbish / remanufacture / recycle / landfill. Computer vision (YOLOv8 + additional condition classifier) + NLP analysis of customer return reason: 88% recall for the "requires remanufacturing" category, 91% accuracy.
Routing through remanufacturing operations
Each returned unit is a graph of operations with branching depending on the state of the components. Stochastic scheduling: MILP with probability weights (the probability that part X will require replacement = 0.4 → the expected operation time). Optimizer: PuLP or Gurobi. On a remanufacturing line with 8 products: throughput increased by 19%, work-in-progress inventory decreased by 28%.
Secondary raw materials management
Material bank and marketplace
Recycled material availability forecasting: return volume × remanufacturing yield rate → recycled material supply. On the other hand: demand forecasting for recycled material buyers. Matching: an auction algorithm (VCG mechanism) for optimally distributing recycled materials among buyers.
Quality grading of recycled materials
Recycled polymer, metal, glass—quality varies by batch. NIR (Near-Infrared) spectroscopy + ML classifier (Random Forest based on spectral features): automatic batch quality classification in 30 seconds vs. 45 minutes of laboratory analysis. Accuracy 94% on 12 polymer classes.
Waste stream optimization
Industrial symbiosis
Industrial symbiosis: waste from one enterprise is raw material for another. ML task: find potential pairs in an industrial cluster. Graph Neural Network (GNN) on a graph of enterprises with attributes (waste type, volume, composition, location, seasonality). Link prediction identifies non-obvious pairs: based on Kalundborg Symbiosis data, seven new potential flows not covered by existing contracts were identified.
Waste composition analysis
Automatic waste sorting: CV system on a conveyor (Greyparrot, Amp Robotics-like architecture). Real-time material classification: plastic by type (PP, PE, PET, PS), metal, cardboard, organics. Accuracy 97% at a conveyor speed of 2 m/s. Waste stream composition data → analytics for optimizing primary material procurement.
Circular design assistance
The LLM agent analyzes the new product's BOM and flags components that hinder recycling, such as incompatible materials, the use of adhesives instead of mechanical fasteners, and missing documentation disassembly. The recyclability score is assessed automatically within the PLM workflow using the Ellen MacArthur Foundation Material Circularity Indicator (MCI) methodology.
Development time: 5–10 months for an end-to-end system. Individual modules (DPP generation, waste sorting CV, reverse logistics forecasting) take 2–4 months each.