2 Literature & resourcesAwesome AI for Industrial Ecology
Literature & resources
A curated reading list with three kinds of entries: AI applied in IE (tagged by IE field and AI technique), research practice (reproducibility, verification, disclosure), and the ML landmarks behind today’s tools. Every entry says in one line why it’s worth your time.
Filter by any combination of tags. Each filtered view has its own URL, so you can share it. Entries live in src/data/literature.yaml. To add one, suggest a paper in a short form, or open a pull request as the contributing guide describes.
- Assessing corporate sustainability with large language models: evidence from Europe
LLM extraction of 501 ESRS indicators from 9,173 corporate reports shows corpus-scale extraction works; the 29% hit rate is non-disclosure, not model failure.
- Evaluation of vision–language models for waste classification: zero- and few-shot with a training-free tip-adapter
Do CLIP-class models sort waste without training data? A measured answer on curated, landfill and multi-material lab images.
- Machine learning and large language models for life cycle inventory compilation: current situation and future perspectives
The most current survey of ML/LLM-assisted LCI compilation; a good entry point as of 2026.
- A call to ensure reproducibility of machine learning applications in industrial ecology
CML survey of ML-in-IE papers; data leakage and poor computational reporting are the two dominant failure modes.
- Reverse Engineering Lab - Data Collection Platform
Open platform for structured product-disassembly data collection (CML): infrastructure for the datasets this field lacks.
- Evaluating large and lightweight vision models for irregular component segmentation in e-waste disassembly
Head-to-head of foundation vs. lightweight segmenters on opened e-waste: what 'SAM needs a class layer' looks like in practice.
- The future of robotic disassembly: a systematic review of techniques and applications in the age of AI
Where robotic disassembly stands: a systematic review of perception, planning, and human-robot collaboration.
- Assessing the performance of 8 AI chatbots in bibliographic reference retrieval
The 2025 update on citation reliability across 8 chatbots: ~40% of retrieved references still erroneous or fabricated.
- SAM 3: segment anything with concepts
Segmentation promptable by concept rather than by click; the class-assignment gap the original SAM left open starts to close here.
- Modeling interior component stocks of UK housing using exterior features and machine learning techniques
Predicts interior walls and rooms from what's visible outside, facade images included: ML as a shortcut around impossible surveys.
- An Expert-grounded benchmark of General Purpose LLMs in LCA
Expert-grounded benchmark of eleven general-purpose LLMs on LCA questions. Experts judged 37% of answers inaccurate or misleading, and citation hallucination reached up to 40%.
- The evolution of research at the intersection of industrial ecology and artificial intelligence
Topic-model map of the IE-and-AI literature: which applications the field has actually been publishing on.
- A bottom-up approach integrating computer vision with material flow analysis to estimate recycling potential
Finds rooftop solar panels in satellite images and feeds a dynamic MFA of their recycling potential; CV and MFA in one pipeline.
- MapAnything: universal feed-forward metric 3D reconstruction
One feed-forward model for metric 3D across many input combinations; the generalist counterpart to VGGT.
- A large language model-based framework to retrieve life cycle inventory and environmental impact data from scientific literature
Fine-tuned LLaMA-2 with RAG extracts LCI data at F1 ≈ 0.82–0.86: good enough to accelerate, not to skip human checks.
- RAISE: a robot-assisted selective disassembly and sorting system for end-of-life phones
A full working cell, from perception to manipulation, for selective phone disassembly; where the integration problems live.
- SINGAPO: single image controlled generation of articulated parts in objects
Articulated parts inferred from one image. Relevant to disassembly structure, but demonstrated only on a few household categories.
- Automated material flow characterization of WEEE in sorting plants using deep learning and regression models
Camera-based deep learning estimates the composition of shredded WEEE on a pilot-scale industrial sorter: CV feeding MFA directly.
- VGGT: visual geometry grounded transformer
Feed-forward 3D from a handful of photos, no per-scene optimisation; multi-view geometry without running a photogrammetry pipeline.
- Reference hallucination score for medical artificial intelligence chatbots
A usable metric for citation hallucination, and evidence that retrieval-backed tools hallucinate far less.
- Depth Pro: sharp monocular metric depth in less than a second
Metric depth from one image in under a second, with boundaries crisp enough to trace a part; the fast end of monocular depth.
- URDFormer: a pipeline for constructing articulated simulation environments from real-world images
Turns a photo into a simulatable articulated model; recovers kinematic joints, not the fastener relations a disassembly plan needs.
- Weighing the global built environment: high-resolution mapping and quantification of material stocks in buildings
Global building material stocks (547 Gt) at 90 m from Earth observation, with an honest 53–200% validation spread.
- Metric3D v2: a versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation
Monocular depth in real metres rather than relative units, plus surface normals; the difference between a nice depth map and a measurement.
- Image-based prediction of residential building attributes with deep learning
Street-view CV predicts floor area (~19% MAPE) and age (~70%) in Toronto, and degrades sharply in other cities; feasibility, not production.
- REFORMS: consensus-based recommendations for machine-learning-based science
A 32-item, field-agnostic reporting checklist to fill in before submitting any ML-based paper.
- Individual model identification of waste digital devices by the combination of CNN-based image recognition and weight information
A CNN and a weight sensor identify individual device models on a sorting line: identity recovery under real EoL conditions.
- Grounding DINO: marrying DINO with grounded pre-training for open-set object detection
Open-set detection driven by a text prompt: turns "find the screws" into boxes without a fixed, pre-trained class list.
- Large language models for life cycle assessments: opportunities, challenges, and risks
A perspective on LLMs in LCA: where they help, where they hallucinate, and why a practitioner should review what they produce.
- SAM 2: segment anything in images and videos
SAM extended to video with memory across frames: segmentation that tracks, still without naming what it found.
- Grounded SAM: assembling open-world models for diverse visual tasks
Grounding DINO and SAM wired together, text prompt in and masks out; the pragmatic open-vocabulary segmentation stack.
- A review of machine learning applications in life cycle assessment studies
Newer ML-in-LCA review; useful as a cross-check on Ghoroghi et al.'s taxonomy of use cases.
- Mitigating grand challenges in life cycle inventory modeling through the applications of large language models
Maps LCI's chronic data problems onto what LLMs can realistically fix, and what they make worse.
- Depth Anything V2
Monocular depth that mostly just works; the geometry-from-one-photo building block. Try it on the demos page.
- Data science applications in circular economy: trends, status, and future
Status check on data science across CE applications, with a sober view of the data bottlenecks.
- Flamingo: environmental impact factor matching for life cycle assessment with zero-shot machine learning
Zero-shot matching of products to impact factors; the Amazon-scale take on automating footprinting.
- High-resolution quantification of building stock using multi-source remote sensing imagery and deep learning
Feature-fusion deep model beats single-source baselines for Beijing's building stock (<14% total error), but still needs ground data.
- AutoPCF: efficient product carbon footprint accounting with large language models
LLM pipeline for product carbon footprints via semantic matching of emission factors; an early working system.
- WinCLIP: zero-/few-shot anomaly classification and segmentation
Zero-shot anomaly detection built on CLIP: flags that something is wrong without defect training data, but not how bad it is.
- Leakage and the reproducibility crisis in machine-learning-based science
The taxonomy of 8 data-leakage types that silently inflate ML results, found in 294 papers across 17 fields.
- 3D Gaussian splatting for real-time radiance field rendering
Radiance fields you can render in real time; where NeRF's quality met a practical speed budget.
- Segment Anything
SAM: promptable segmentation of anything. Try it on the demos page, then note it can't tell you what the segments are.
- Reproducibility in machine-learning-driven research
Quick map of why 'we used ML' so often means 'nobody can rerun it': barriers and the practices that fix them.
- Product label identification with OCR for model-specific reuse, repair, and recycling: A case study for washing machines
Deployed label-OCR pipeline for e-waste identification; reads 95.7% of labels in repair shops but 39.7% at the recycler.
- A systematic review of Green AI
What's known about making AI itself less resource-hungry: an IE question applied to our own tools.
- Fabrication and errors in the bibliographic citations generated by ChatGPT
Measured, not anecdotal: 55% of GPT-3.5 and 18% of GPT-4 citations were fabricated; many real ones still had errors.
- Applications of convolutional neural networks for intelligent waste identification and recycling: A review
Systematic review of CNN waste-identification work: datasets, architectures, and the recurring benchmark gaps.
- A survey of large language models
The reference survey when you need the full LLM landscape (training, capabilities, evaluation) in one document.
- AnomalyCLIP: object-agnostic prompt learning for zero-shot anomaly detection
Prompts learned to be object-agnostic, so zero-shot anomaly detection travels across product types instead of being tuned per object.
- The future of artificial intelligence in the context of industrial ecology
The JIE perspective on what AI can and cannot do for IE; a good starting point for this whole resource.
- Advances in application of machine learning to life cycle assessment: a literature review
Broad review of where ML enters LCA workflows: impact prediction, inventory estimation, interpretation.
- Computer vision for solid waste sorting: A critical review of academic research
The critical review of CV-for-waste-sorting research, including why lab accuracies don't survive the sorting plant.
- Deep learning-based waste detection in natural and urban environments
Waste detection benchmarks and datasets in one place, plus semi-supervised tricks for the label-scarcity problem.
- Training language models to follow instructions with human feedback
InstructGPT/RLHF: why chat models answer questions instead of continuing your sentence, and where their politeness bias comes from.
- A review of inventory modeling methods for missing data in life cycle assessment
The gap-filling problem stated properly: what's missing in LCI data and which estimation methods (including ML) apply.
- A scalable data collection, characterization, and accounting framework for urban material stocks
Mobile sensing and CV count doors and windows across a whole neighborhood at accuracy comparable to manual surveys.
- High-resolution maps of material stocks in buildings and infrastructures in Austria and Germany
National material stock maps at 10 m from free satellite data and OSM, no cadastre needed; ~5 Gt Austria, ~38 Gt Germany.
- Highly accurate protein structure prediction with AlphaFold
The clearest case of deep learning solving a real scientific problem, and the bar for 'AI for science' claims.
- Estimation and mapping of the material stocks of buildings of Europe: a novel nighttime lights-based approach
Night-lights radiance as a continent-scale proxy for built-up volume: MFA without a single site visit.
- Learning transferable visual models from natural language supervision
CLIP: images and text in one embedding space; the zero-shot recognition used in this site's own demo.
- Disassembly-based bill of materials data for consumer electronic products
95 products physically disassembled into validated BoM data: the ground truth CV-based approaches need to beat.
- Language models are few-shot learners
GPT-3: scale alone produced in-context learning. The paper that made 'prompting' a verb.
- An image is worth 16x16 words: transformers for image recognition at scale
ViT: transformers displace convolutions in vision, the reason 'foundation model' now includes images.
- Machine learning based modeling of households: a regionalized bottom-up approach to investigate consumption-induced environmental impacts
ML clustering of household data to model consumption footprints bottom-up across a whole country.
- Denoising diffusion probabilistic models
The diffusion formulation behind modern image generation: noise in, image out, one denoising step at a time.
- Scaling laws for neural language models
Why bigger kept getting better: the empirical laws that justified training ever-larger models.
- NeRF: representing scenes as neural radiance fields for view synthesis
The paper that made photorealistic 3D-from-photos routine. Slow to fit per scene, and the ancestor of everything that followed.
- Quantifying the carbon emissions of machine learning
The standard starting point (and calculator) for reporting the footprint of your own model training.
- Rapid life-cycle impact screening using artificial neural networks
Early proof that ANNs can screen chemical impacts orders of magnitude faster than full characterization.
- Attention is all you need
The transformer architecture behind every modern LLM; worth knowing what 'attention' does.
- Material recognition in the wild with the Materials in Context Database
MINC: 23 material classes on real in-the-wild photos, and still the reference point for recognising what a surface is made of.
- Deep residual learning for image recognition
ResNet: the trick (skip connections) that made very deep networks trainable; still inside most vision backbones.
- You only look once: unified, real-time object detection
Detection as a single regression pass; the reason real-time object detection is cheap enough to run on a sorting line.
- Big data and industrial ecology
The early (2015) case for data-intensive IE; useful context for everything newer on this list.
- ImageNet: a large-scale hierarchical image database
The dataset that kicked off the deep-learning era, and the reason benchmarks steer the field.