1 Map of AI applicationsAwesome AI for Industrial Ecology
Map of AI applications in IE
This map shows where AI is applied in industrial ecology, along two dimensions: the IE field (LCA, MFA, waste & circularity, …) and the AI technique (LLMs, computer vision, tabular ML, …). Every entry links to the literature list pre-filtered to the same tags, so the map is also an index into the reading list.
What works today, per field
Section titled “What works today, per field”One section per field: a paragraph on what works today, then the field’s applications, each with a maturity call. Open an application for its description and the papers behind it, or see every paper in the literature list.
LCA & footprinting
Section titled “LCA & footprinting”Two things demonstrably work. The first is tabular ML for rapid impact screening and inventory gap-filling, validated repeatedly since 2017. The second is LLM extraction of inventory data from literature and reports: working systems reach an F1 score of 0.82–0.86 (1 would mean every value found and none wrong) (Kumar et al., 2025), with a person checking the values. The papers here use LLMs to extract inventory data, not to generate it (Preuss et al., 2024). A 2025 benchmark found 37% of general-purpose LLM answers to LCA questions inaccurate or misleading, with hallucinated-citation rates up to 40% (Donaldson et al., 2025). Automated footprinting works for products with a clear match in the emission-factor database. In one evaluation, a few wrong matches caused most of the error (Balaji et al., 2023). Papers →
6 applications in LCA & footprinting
Automated product carbon footprintingWorks for products with a clear database match. In one evaluation, a few wrong matches caused most of the error (MAPE 18% with a product-code filter, 144% without). Audit trails remain the weak point.
Match products to emission factors semantically (zero-shot or LLM-driven) to footprint whole product catalogs.
LCA & footprintingLLMs & agentsNLP & information extractionpapers (2) →Household consumption-footprint modelingDemonstrated at national scale; depends on rich household microdata that most countries don't have.
Cluster household microdata and link archetypes to environmental extensions for regionalized, bottom-up consumption footprints.
LCA & footprintingTabular ML & regressionpapers (1) →Inventory gap-filling & estimationActive research area; method choice matters more than model class, and extrapolation beyond training data stays risky.
Estimate missing life-cycle inventory values from similar processes and product attributes instead of leaving gaps or copying proxies.
LCA & footprintingTabular ML & regressionpapers (2) →LLM extraction of LCI data from literatureWorking systems reach F1 ≈ 0.82–0.86. The papers here use LLMs to extract values for a person to check, not to generate missing data.
Fine-tuned or retrieval-augmented LLMs pull inventory and impact data out of papers, reports, and MSDS into structured databases.
LCA & footprintingLLMs & agentsNLP & information extractionpapers (3) →Rapid impact screening with tabular MLRepeatedly demonstrated since 2017; screening-grade accuracy, not a replacement for characterization.
Train regression/ANN models on existing characterization results to screen chemicals or products orders of magnitude faster than full assessment.
LCA & footprintingTabular ML & regressionpapers (3) →Sustainability-indicator extraction from corporate reportsCorpus-scale extraction validated in 2026; the binding constraint is what companies disclose, not the model.
Extract quantitative ESG/ESRS indicators from thousands of corporate reports at once, turning piles of PDFs into panel data.
LCA & footprintingNLP & information extractionLLMs & agentspapers (1) →
MFA & stocks and flows
Section titled “MFA & stocks and flows”Satellite stock mapping is the furthest along here. It works at national scale (10 m) (Haberl et al., 2021) from free data, and a global estimate (90 m, 547 Gt in buildings) (Haberl et al., 2024) exists. Against independent studies, results range from 53% to 200%. Neither side is ground truth, so treat that spread as the uncertainty. Night-lights proxies (Peled & Fishman, 2021) and deep multi-source fusion (Bao et al., 2023) improve on single sources but still need ground data. Street-view CV for building attributes is demonstrated feasibility: good in the training city, sharply worse elsewhere (Huang et al., 2024). Papers →
5 applications in MFA & stocks and flows
Building attributes from street-level imageryFeasibility shown (≈19% MAPE floor area, ~70% age accuracy in-city); transfer between cities degrades sharply.
Predict floor area, age, and component counts (windows, doors) from street-view photos to feed urban stock models.
MFA & stocks and flowsComputer visionDetection & segmentationpapers (2) →Interior component stocks from exterior featuresOne study, 2,000 UK houses; useful priors, wide uncertainty per building.
Predict interior wall length and room counts from exterior features and facade images, as input to building stock models: a shortcut around impossible surveys.
MFA & stocks and flowsUrban mining & secondary resourcesTabular ML & regressionComputer visionpapers (1) →Material stock mapping from satellite dataNational 10 m maps and a global 90 m estimate exist. Compared with independent studies, results range from 53% to 200%; neither side is ground truth, so treat that spread as the uncertainty.
Combine Earth-observation imagery, OSM, and material intensities to map building/infrastructure stocks without cadastral data.
MFA & stocks and flowsRemote sensingpapers (3) →Night-lights & multi-source stock estimationContinent-scale proxy validated; deep fusion improves accuracy but still needs ground data for calibration.
Use nighttime-light radiance, alone or fused with optical imagery by deep models, as a proxy for built-up volume and stocks.
MFA & stocks and flowsRemote sensingpapers (2) →Sensor-based material flow characterization in sorting plantsTested on an industrial sorter at pilot scale with shredded WEEE (mAP50 ≈ 0.99, composition within about 5%), on recorded data. No plant runs it yet.
Cameras on sorting lines classify waste particles and estimate the composition of the flow, which could feed MFA with plant data.
Waste, EoL & circularityMFA & stocks and flowsComputer visionDetection & segmentationpapers (1) →
Waste, EoL & circularity
Section titled “Waste, EoL & circularity”The corner with the most publications. Detecting and counting waste and components is accurate on the benchmarks a model was tuned for, and untested on field photos (Majchrowska et al., 2022). On a pilot-scale industrial sorter, cameras estimated the composition of shredded e-waste (WEEE) from recorded data; no plant runs it yet (Vogelgesang et al., 2025). Reading product labels with text recognition (OCR) is deployed in pilots, where readability collapses from 96% in repair shops to 40% at the recycler (Sterkens et al., 2023). Worn, dirty and damaged products are the recurring failure mode. Foundation models (SAM, CLIP) demo impressively. They still need a class-assignment layer and sorting-grade evaluation before they count as methods. Robotic disassembly works for narrow product families (Ameur et al., 2025). Papers → · Full depth in the CV taxonomy worked example.
7 applications in Waste, EoL & circularity
Component detection & counting in opened productsAccurate on the benchmarks it was tuned for, untested on field photos. Foundation segmenters (SAM) still need a class-assignment layer.
Detect, segment, and count components inside opened devices to build bills of components for disassembly and recycling.
Waste, EoL & circularityComputer visionDetection & segmentationpapers (2) →demo →Geometry & quantity from single imagesGives approximate geometry from one photo. Accuracy on field photos of real products is untested; reflective, transparent, and thin parts fail.
Monocular depth and feed-forward 3D reconstruction recover error-bounded product geometry from ordinary photos.
Waste, EoL & circularityDepth & 3D geometryComputer visionpapers (1) →demo →Product identification from labels (OCR + retrieval)Deployed in pilots; reads 95.7% of labels in repair settings but 39.7% at the recycler; label wear is the limit.
Read model/serial text off product labels and match it against a reference database to identify products at end-of-life.
Waste, EoL & circularityUrban mining & secondary resourcesComputer visionNLP & information extractionpapers (2) →Sensor-based material flow characterization in sorting plantsTested on an industrial sorter at pilot scale with shredded WEEE (mAP50 ≈ 0.99, composition within about 5%), on recorded data. No plant runs it yet.
Cameras on sorting lines classify waste particles and estimate the composition of the flow, which could feed MFA with plant data.
Waste, EoL & circularityMFA & stocks and flowsComputer visionDetection & segmentationpapers (1) →Vision-guided robotic disassemblyWorking cells exist for narrow product families (phones); generalization across products is the open problem.
Perception + planning pipelines locate fasteners and components to disassemble products robotically for reuse and recycling.
Waste, EoL & circularityComputer visionDetection & segmentationpapers (2) →Waste image classification & detectionThe most-published application in the field; lab accuracies are high, but performance under plant conditions is the real test.
CNN/transformer classifiers recognize waste categories in images, from litter surveys to conveyor streams.
Waste, EoL & circularityComputer visionpapers (3) →Zero-shot recognition with vision-language modelsWithout training, CLIP sorts single-material waste photos well (about 90% on a curated set, 79% on landfill photos) but drops to 40% when items carry several materials. Not yet sorting-grade.
CLIP-class models score user-chosen labels against images: recognition without task-specific training data.
Waste, EoL & circularityComputer visionFoundation models & transformerspapers (2) →demo →
Urban mining & secondary resources
Section titled “Urban mining & secondary resources”Wanted A paragraph on what works here, from someone who knows the field. Papers →
2 applications in Urban mining & secondary resources
Interior component stocks from exterior featuresOne study, 2,000 UK houses; useful priors, wide uncertainty per building.
Predict interior wall length and room counts from exterior features and facade images, as input to building stock models: a shortcut around impossible surveys.
MFA & stocks and flowsUrban mining & secondary resourcesTabular ML & regressionComputer visionpapers (1) →Product identification from labels (OCR + retrieval)Deployed in pilots; reads 95.7% of labels in repair settings but 39.7% at the recycler; label wear is the limit.
Read model/serial text off product labels and match it against a reference database to identify products at end-of-life.
Waste, EoL & circularityUrban mining & secondary resourcesComputer visionNLP & information extractionpapers (2) →
Explore by technique
Section titled “Explore by technique”The same applications as a grid of IE field × AI technique, or grouped by technique:
Table 1.1 Applications by AI technique and IE field
| LCA & footprinting | MFA & stocks and flows | Waste, EoL & circularity | Urban mining & secondary resources | |
|---|---|---|---|---|
| LLMs & agents | 3 | No application yet: MFA & stocks and flows with LLMs & agents | No application yet: Waste, EoL & circularity with LLMs & agents | No application yet: Urban mining & secondary resources with LLMs & agents |
| NLP & information extraction | 3 | No application yet: MFA & stocks and flows with NLP & information extraction | 1 | 1 |
| Computer vision | No application yet: LCA & footprinting with Computer vision | 3 | 7 | 2 |
| Detection & segmentation | No application yet: LCA & footprinting with Detection & segmentation | 2 | 3 | No application yet: Urban mining & secondary resources with Detection & segmentation |
| Depth & 3D geometry | No application yet: LCA & footprinting with Depth & 3D geometry | No application yet: MFA & stocks and flows with Depth & 3D geometry | 1 | No application yet: Urban mining & secondary resources with Depth & 3D geometry |
| Remote sensing | No application yet: LCA & footprinting with Remote sensing | 2 | No application yet: Waste, EoL & circularity with Remote sensing | No application yet: Urban mining & secondary resources with Remote sensing |
| Tabular ML & regression | 3 | 1 | No application yet: Waste, EoL & circularity with Tabular ML & regression | 1 |
| Foundation models & transformers | No application yet: LCA & footprinting with Foundation models & transformers | No application yet: MFA & stocks and flows with Foundation models & transformers | 1 | No application yet: Urban mining & secondary resources with Foundation models & transformers |
Shade shows the number of applications: 1, 2, or 3 or more. A dashed cell has none yet. Source: src/data/applications.yaml.
Not mapped yet
Section titled “Not mapped yet”The gaps in this map matter as much as the filled cells, so we name them here instead of leaving them implicit. Two kinds of gap: IE fields with no vetted entries, and AI techniques with no tag in the vocabulary yet.
A caveat on how to read this: none of it claims the work doesn’t exist. It claims we have not vetted any of it. Our sweep either found nothing that survived verification, or never looked. Each line is the easiest possible contribution: one paper, one entry. Its link opens a short form on GitHub with the gap filled in.
IE fields with no entries
Section titled “IE fields with no entries”- Wanted Chemicals, toxicity & QSAR. One LCA entry screens chemicals’ impacts from their molecular structure. Nothing yet on toxicity characterisation factors for the tens of thousands of substances nobody has tested. Plausibly the largest single hole in this map. Suggest a paper on Chemicals, toxicity & QSAR
- Wanted Energy systems & industrial decarbonisation. Marginal versus average grid factors, temporal accounting, process and heat-integration optimisation. Suggest a paper on Energy systems & industrial decarbonisation
- Wanted Industrial symbiosis. Matching one plant’s waste to another’s input is, in part, a recommender problem. Suggest a paper on Industrial symbiosis
- Wanted Supply chains, criticality & resource security. Supply risk, critical raw materials, trade-network disruption. Suggest a paper on Supply chains, criticality & resource security
- Wanted Agriculture, food & land use. The field our remote-sensing tag most obviously serves, with nothing behind it. Suggest a paper on Agriculture, food & land use
- Wanted Product lifetime, reuse & repair. Survival analysis and failure prediction; feeds both MFA stock models and the EoL corner. Suggest a paper on Product lifetime, reuse & repair
- Wanted Water. Water footprints, scarcity characterisation, and hydrological modelling. Suggest a paper on Water
- Wanted Input–output analysis, forecasting & scenario modelling. Economy-wide footprints and what-if scenarios. Suggest a paper on Input–output analysis, forecasting & scenario modelling
AI techniques with no tag
Section titled “AI techniques with no tag”Tags are added together with their first entry, so these stay out of the vocabulary until someone brings a paper that earns one.
- Wanted Uncertainty quantification & Bayesian methods. Uncertainty is this discipline’s central methodological concern, and the map has no column for it. Conformal prediction counts too. Suggest a paper on Uncertainty quantification & Bayesian methods
- Wanted Graph neural networks & network analysis. Supply chains, IO tables, material-flow networks, molecular graphs: most IE data is a network, and nothing here treats it as one. Suggest a paper on Graph neural networks & network analysis
- Wanted Surrogate models & emulators. Physics-informed ones included. This is how scenario work gets done with ML in practice, not just in principle. Suggest a paper on Surrogate models & emulators
- Wanted Causal inference. Rebound effects and policy attribution are causal claims, usually made with correlational tools. Suggest a paper on Causal inference
- Wanted Reinforcement learning. Disassembly sequence planning and process control. The EoL paragraph above already discusses robotic disassembly with no technique tag behind it. Suggest a paper on Reinforcement learning
- Wanted Knowledge graphs & ontologies. LCI interoperability and semantic flow matching, distinct from NLP extraction. Suggest a paper on Knowledge graphs & ontologies
- Wanted Recommender & matching systems. The method behind industrial symbiosis, and behind matching products to emission factors. Suggest a paper on Recommender & matching systems
- Wanted Time-series & forecasting. Distinct from tabular ML. Suggest a paper on Time-series & forecasting
- Wanted Federated & privacy-preserving learning. Local AI & hardware argues for keeping sensitive data in-house; this is how institutions learn from data they are never going to pool. Suggest a paper on Federated & privacy-preserving learning
- Wanted Explainability. Whether someone can interrogate a claim from a model-based assessment. Suggest a paper on Explainability
Where the gaps meet
Section titled “Where the gaps meet”Several of these are not separate holes but one cell each: chemicals × graph neural networks, scenarios × surrogates, symbiosis × matching, EoL × reinforcement learning. Those four are the most concrete places to start, because each names both a field and a method rather than a vague area.
Deeper dives
Section titled “Deeper dives”- Worked example: CV for product-scale data collection: one corner of this map (computer vision at end-of-life) taken to full depth, with per-cell maturity verdicts. The pattern to copy when your own corner deserves more than a map entry.