Artificial intelligence in non-ferrous metals and recycling
Three reference applications, selected according to criteria K1–K5 from cases described in public sources from 2021–2026. The ranking reflects the state of publication, not the full picture of deployments: plants that do not publish results remain outside the review.
Al electrolysis · EGA · P2/N22–3 days → < 1 shifttime to detect anode disturbances relative to previous practice; 2,592 pots, 30% kept as a reference for assessing benefits
Processing · review 2021–20260published AI results with a control group for a named rolling, extrusion, wire rod, tube or foil line
Metals and metal products · Eurostat · 202514.3% EU · 7.2% PLshare of enterprises (10+ persons employed) using at least one AI technology; the data also cover steel
Prepared by: Economic Chamber of Non-Ferrous Metals and Recycling · as of 25.10.2026 · reviewed every 6 months, next review by 30.04.2027 · prepared with the use of AI tools
Summary
Five theses
The best-documented AI case concerns primary smelting. In the review, only EGA describes a result measured against a reference group of pots [1], in a programme assessed externally [2]. This does not indicate that AI is of lesser value in processing, but that publications from that area are lacking.
In recycling, alloy composition is measured by a sensor, and deep learning decides where shape and context matter. Sorting of 5xxx/6xxx alloys by LIBS operates at a scale of thousands of tonnes without deep learning – relying on chemometric classification and rules [6]. No one has yet measured the contribution of AI to the output of sensor cascades [9][10].
Machine vision at the line is the most frequently described class of AI in the review, but without published measured results. According to the platform vendor, EGA reports inspecting 100% of carbon anodes [14]; Jiangxi Copper, Ningbo Jintian and Hindalco have deployed vision inspection of cathodes and strip [15][16][41].
A credible result requires a measurement design, not just a model. At EGA: advisory mode, a reference group of pots and confirmation of the benefits by the finance department before the deployment was extended [1].
The year 2027 mandates data that double as input for AI. The battery passport (18.02.2027), the restriction on waste exports outside the OECD (21.05.2027), the first CBAM declaration (30.09.2027) and the targets for recovery of metals from batteries (31.12.2027) require a mass balance and batch traceability [24][25][38].
Methodology
Scope, evidence assessment and selection of the TOP 3
Scope: from concentrate and charge to product, plus the scrap loop – smelting, refining, processing and recycling of non-ferrous metals. Ore extraction and beneficiation belong to mining and remain background.
Review protocol
Period: 01.2021–10.2026; last search 25.10.2026. Sources: companies’ annual, sustainability and stock-exchange reports (including SEC and CNINFO), company press releases, ICSOBA and International Copper Conference materials, peer-reviewed journals, CORDIS, MIIT lists, the WEF Global Lighthouse Network report, EUR-Lex, ISAP, Eurostat, EuroHPC and the Funding & Tenders portal. Languages: English, Chinese, German, Polish, Russian, Spanish.
Procedure: four thematic searches (recycling, primary production, processing, 2027 directions), followed by six independent fact audits – each claim was checked directly against the source document – as well as a technical audit and a scientific review. 36 cases were assessed: 21 were included (the TOP 1–3 and “Beyond TOP 3” tables) and 15 excluded (Table 2); in addition, 3 figures known only from secondary reports were rejected.
The searches and audits were performed with the use of AI tools, without pre-registration of the protocol. The full texts of TMS Light Metals – the main publication venue for work on AI in aluminium electrolysis – were not available; EGA’s 2024 papers are known only from the index.
Evidence assessment in two dimensions
Measurement design (P)
P1
control group with described selection
P2
reference group declared, selection not described
P3
“before and after” comparison or comparison with previous practice
P4
no attribution of the effect, or no figures
L
laboratory result or preprint – does not meet K1
Source independence (N)
N1
peer-reviewed publication or third-party assessment with a disclosed method
N2
conference paper, report or company document
N3
technology vendor or consultant
†
state news agency release as a substitute source for Chinese companies
Rules
A figure without a defined measure, or reported only in the press, does not raise the rating. A primary source is a document of a company, an authority or a legal act, or a publication by the authors of the deployment; the trade press was used only to locate sources.
P2/N2P3/N1P4/N2P4/N3L/N1
Selection criteria
Prerequisite W0: the source explicitly describes an AI or machine-learning method. Each criterion was scored on a 0–2 scale with equal weight; ties are broken by the higher K2 score, and then by K3.
K1Industrial scale0 – laboratory or single pilot · 1 – one plant · 2 – multiple plants or more than 1,000 units
K2Attribution of effect0 – P4 · 1 – P2 or P3 · 2 – P1
K3Source0 – press or vendor only · 1 – company report or state agency · 2 – publication by the authors of the deployment or third-party assessment
K4Transferability0 – one-off solution · 1 – described method · 2 – method commercially available or replicated at another plant
K5Value-chain relevance0 – marginal · 1 – quality or cost of one link · 2 – energy, emissions or metal circularity at industry scale
Table 1. Scoring matrix of application classes
Application class
W0
K1
K2
K3
K4
K5
Total
Result
Al electrolysis – soft sensors and disturbance detection
yes
2
1
2
2
2
9
TOP 1
Scrap sorting – sensor cascade with AI
yes
2
0
1
2
2
7
TOP 2
Vision and prediction at the line
yes
2
0
1
2
1
6
TOP 3
Short-circuit detection in copper electrorefining
no
2
0
1
1
2
6
not qualified – sources do not indicate AI
Flotation – hybrid model and MPC
yes
2
1
2
1
1
7
out of scope (mining) – background
Industry language models and agents
yes
1
0
1
1
1
4
2027 direction
TOP 3 has the lowest level of evidence of the three highlighted classes: its position results from the number of independent deployments (K1) and the availability of the technology (K4), not from the strength of evidence. Short-circuit detection in electrorefining has the same total (6) but does not meet prerequisite W0 – the sources do not identify an AI method – and is therefore not ranked.
Figure 1. Non-ferrous metals value chain and where the TOP 1–3 applications sit
Own work, illustrative diagram. Green outline: links with TOP 1–3 applications; green arrows: secondary loop. Labels indicate the method classes most often described in the 21 included cases.
TOP 1
Aluminium electrolysis
Soft sensor and disturbance detection in electrolytic pots
A model trained on hourly pot data supplements the infrequent bath temperature measurement and flags anodes requiring intervention, shortening the duration of a disturbance that reduces current efficiency.
Plant
Scope
Method
Result
Rating
EGA – Jebel Ali and Al Taweelah (UAE)
1,572 pots at Jebel Ali and 1,020 selected pots at Al Taweelah; 30% kept as a reference for assessing benefits
Random Forest (classification and regression), Bayesian tuning; more than 30 parameters aggregated hourly; soft sensor for bath temperature (thermocouple measurement every 48 h), mean error 4.4 °C
detection of anode disturbances in less than one shift instead of 2–3 days (relative to previous practice); “audited” gains of 3,050–9,070 t of metal – a range depending on pot technology, with no period and no auditor named
RBF network on individual anode currents for detecting a poorly set anode (2018); low-voltage anode effect model calibrated with CF₄ measurements (2020)
no plant results; low-voltage anode effects account for up to 50% of PFC emissions in some pot designs
The values from [1] require clarification: recall of 75% and precision of 35% come from retrospective validation on three months of data for 57 pots selected by stratified sampling, not from prospective measurement; the publication does not describe how the reference group was selected. The Al Taweelah smelter was damaged in March 2026; according to Reuters (26.08.2026), 315 of 1,262 pots had been restarted, with full production expected in Q1 2027 [31] – the results in the table predate this event. The exclusion of the Chalco Guizhou case is justified in Table 2.
Figure 2. Electrolytic pot with a data layer and an advisory loop
Own work based on [1], illustrative diagram, not to scale. Bath temperature indicative, depending on composition (excess AlF₃). The loop is closed by the operator – the model does not control the pot directly.
For management
Value mechanism: an anode disturbance detected days earlier lasts for a shorter time, and every hour of disturbance reduces current efficiency; the effect adds up across thousands of pots.
At an energy consumption of approx. 13,500 kWh/t (DC), a 1 percentage point increase in current efficiency corresponds to approx. 140 kWh/t, and a 10 mV voltage reduction to approx. 32 kWh/t.
According to the company, EGA’s Industry 4.0 programme has delivered approx. USD 100 million in benefits since 2021; the figure relates to the whole programme, not to a single model [7]. The sources disclose neither the investment nor the payback period.
Entry condition: a historian with several years of hourly data, and process engineers who verify the alerts.
For engineers
Phenomenon
anode spike – a protrusion on the underside of the anode that approaches or touches the metal: the anode–cathode distance decreases, anode current and temperature rise, and current efficiency falls. An anode effect involves a drop in Al₂O₃ concentration below approx. 1.5–2 wt%, a voltage jump (or a low-voltage signature) and emission of CF₄ and C₂F₆ – a different problem, with different models.
Data
pot voltage, anode currents (voltage drop across the anode rod), feeding, operating events; labels from the disturbance history; thermocouple reading every 48 h as the target variable of the soft sensor. The 4.4 °C error should be assessed against bath superheat (typically 5–15 K), which is sometimes the appropriate target quantity.
Model
tree ensemble (Random Forest) – stable with a small number of events and easy to explain; Bayesian tuning
Metrics
recall 75% at precision 35% (F1 ≈ 0.48) – on average approx. 2 false alarms per real disturbance; according to the authors, probing the anode after an alarm is cheap; the cost ratio is not given. Report the inspection workload per shift, the distribution of alarm lead time and PR-AUC, not “accuracy”.
Conversion
WDC [kWh/t] ≈ 2,980 · U [V] / η; assuming η ≈ 94%, U ≈ 4.26 V; DC energy at the pot busbars, AC higher by approx. 2–3% due to rectifier losses
Validation
time-based data split (no random split – information leakage); out-of-time test separately for each pot technology; shadow mode before advisory mode; KPIs fixed in advance; reference group with described selection
Integration
control-room dashboard, voice alert and SMS, work order for the potline crew or in the CMMS, link to the anode change schedule; loop closure only after the advisory period
Latency
decisions on a timescale of minutes to an hour – central inference is sufficient
Reproducibility
number of labelled events and class balance, data period and split, retraining frequency, method of selecting reference pots
Limitations. The metal gain in [1] is given as a range across pot technologies, without a period – it must not be annualised or summed. The results relate to 230–465 kA pots; transfer to another technology requires retraining and its own reference group.
TOP 2
Recycling
Sensor cascade with deep learning in aluminium scrap sorting
X-ray transmission separates materials by effective atomic number, a neural network working on images rejects castings and contaminants, and laser spectroscopy assigns each piece to an alloy series. Only the combination of these stages yields feedstock for sheet and profiles.
Entity
Role of AI
Scale and result
Rating
Novelis – Berea (USA) with EverestLabs
AI and robotics for sorting, tracking and analysing scrap (RecycleOS); pilot
since 09.2024, more than 2 million contaminants identified and more than 1 million removed – a figure without a denominator (tonnage, time, change in purity) and not a measure of effectiveness; recycled content in Novelis flat-rolled products 63% (FY25)
deep learning on RGB images downstream of an XRT separator
up to 2,000 ejections per minute; removes from the wrought fraction the low-alloy castings that XRT cannot distinguish from it; for used beverage cans (UBC), up to 33× the throughput of manual sorting, purity at least 98%
No source measures the gain in purity or yield contributed by the deep-learning stage relative to an XRT + LIBS cascade without AI; the contribution of AI to the cascade result remains undetermined. The sensor layer without deep learning operates at industrial scale: Constellium at Neuf-Brisach remelted more than 6,000 t of scrap sorted by LIBS into the 5xxx and 6xxx families at a purity above 95% (as of Q4 2024) [6]; Hydro sorts with HySort technology at sites including Dormagen (36,000 t per year) and Nowa Sól [13].
Figure 3. Aluminium scrap sorting cascade: XRT, deep learning, LIBS
Own work based on [9], [10], [12], illustrative diagram. In practice, each station is a separate sorter with in-flight ejection beyond the end of the belt; the common belt is a simplification. Station order according to the TOMRA and Novelis test; fraction composition indicative.
For management
The value lies in the difference between the price of mixed scrap and that of feedstock of known alloy: sorted scrap can return to the same product instead of going into castings.
A product with a guaranteed share of post-consumer scrap (e.g. Hydro CIRCAL, at least 75%) requires alloy sorting as a condition of supply [13].
When procuring an AI cascade, require measurement of purity and yield per fraction, broken down by the contribution of each station – vendors do not publish this today.
The sources disclose neither the investment nor the payback period.
For engineers
Preparation
shredding, magnetic and eddy-current separation, screening; singulation of pieces – a precondition for measuring a single object
Decision
in the cascade, each station decides separately; object-level fusion (LIBS + RGB-D at one station, [12]) is a laboratory result
Sensors
XRT does not recognise the alloy series; XRF sees Mg and Si poorly; LIBS measures Mg, Si and Li, but only the surface layer – cleaning pulses remove coatings
Labels
golden set of pieces with OES (spark) analysis or handheld LIBS; XRF only for Cu, Zn, Fe, Mn; inter-annotator agreement; revalidation after a change of scrap supplier; ultimately, liquid-metal analysis at the melt
Latency
example: a 3 m/s belt and 0.4 m to the ejector give approx. 130 ms for measurement, inference and ejection; inference of the order of 10–20 ms
KPI
fraction purity, yield (recovery of metal into the correct fraction), t/h, share of pieces without a decision
Reproducibility
composition and origin of the feed, particle size, belt speed, LIBS frequency, purity measurement method
Limitations. The TOMRA figures relate to tests and are a vendor claim. The Novelis result relates to a pilot at one plant. No source discloses the edge computing hardware.
TOP 3
Smelting, refining and processing
Machine vision and prediction at the line – from anode to strip
A camera with a model at the line replaces sampling inspection with inspection of every piece, and a predictive model indicates the line condition before a disturbance occurs. This is the most frequently described class of AI in the review – and the least documented with figures.
Plant
Stage
Application
Result
Rating
EGA (UAE)
carbon anodes (Al electrolysis)
edge AI vision, network trained on 100,000 images
according to the platform vendor: inspection of 100% of anodes instead of 1.5–2.5%, response 10–13 times faster
more than 50 AI and ML models in the year; reliability models on more than 700 critical assets; vision inspection of cathode surfaces at Birla Copper Dahej
Own work – reference architecture derived from cases [14]–[17] and [41], illustrative diagram; does not represent a specific plant. Resolution and latency are chosen according to line speed and the smallest critical defect in the customer specification.
For management
Inspection of every piece changes the economics of customer claims: a defect is assigned to a coil and position, and the decision is made before shipment.
No processor in the primary sources has published a measured effect of AI for a named line.
When procuring a system, require metrics per defect class and acceptance on the plant’s own data (FAT and SAT), not general claims.
The sources disclose neither the investment nor the payback period.
For engineers
Acquisition
line-scan cameras on the top and bottom; pixel no larger than 1/3 of the smallest critical defect; for bright Al and Cu surfaces, diffuse or coaxial lighting; emulsion and oil cause false alarms; on hot strip, protective housings and cooling
Latency
camera line rate = speed / pixel size, e.g. 20 m/s at 0.2 mm gives 100 kHz; marking in under 1 s; setpoint adjustments in minutes and hours
Model
defect detection and classification; supplement rare classes with synthetic images and active labelling; an operator image-review station as a source of labels
KPI
recall for critical defects at a given number of false alarms per kilometre of strip; 100% surface coverage
Integration
defect metadata (coil ID, position, class, confidence) via OPC UA (Machine Vision, OPC 40100) or MQTT; images in object storage; position from level 2 automation; genealogy according to ISA-95 (B2MML); on model failure, the section is flagged in the MES as “not inspected”
Prediction
for wire rod lines and presses – abnormal-state models on process data; a setpoint advisor as in the Constellium Valais project
Limitations. The EGA anode inspection metrics are reported by the cloud platform vendor. The Ningbo Jintian figures (yield +2%) come solely from a press interview and were not included. The Constellium Valais project is a preprint without confirmed production status.
Beyond TOP 3
Other cases and excluded cases
Area
Case
Result
Rating
Flotation (background)
Nornickel – Talnakh: hybrid model (physics + machine learning), model predictive control (MPC)
by the end of 2024, overall recovery Ni +0.36% rel., Cu +0.16% rel. (to bulk concentrate); the source does not give the baseline
rule-based system, “before and after” comparison, physically impossible figures – e.g. 2.2 t Al per day from a 230 kA pot, which at 100% current efficiency yields approx. 1.85 t [5]
K3
Codelco – Chuquicamata, Ministro Hales
the source describes “advanced analytics and mathematical algorithms”, not AI
W0
Freeport – leaching (214 million lb Cu per year)
the company attributes the result to “data analytics”, not AI
W0
Tongling Jinguan – short circuits in copper electrorefining; ESCD system
thermal imaging and digital twin; sources do not indicate AI [32][33]
W0
Constellium (LIBS), Hydro HySort, Gerhard Lang, STEMIN
sensor-based sorting with no indication of AI – included as context in TOP 2
W0
Aurubis – sampling in Lünen and Hamburg; sensors in Olen
automation, robotics and predictive maintenance with no indication of AI
W0
Boliden Kokkola – anode cleaning
machine vision with no indication of AI
W0
Hindalco – billet butt detection on the press
vibration sensing and a PLC, not AI
W0
Apple – Cora line and disassembly robots
the report does not indicate AI; consumer electronics
W0
Siemens Industrial Copilot
no named non-ferrous metals customer
K1
Figures only from secondary reports: Freeport Bagdad (+10%, consultant), STEINERT IOI (+40%, press), Ningbo Jintian (+2%, interview)
no primary source for the figure
K3
China: the MIIT list of “excellence-level” smart factories (2024) includes 10 non-ferrous metals plants, among them the Yunnan Wenshan aluminium electrolysis plant (“AI-based control”), the Jiangxi Copper smelter and Ningbo Jintian (“AI + big data”) [16].
Architecture
A common framework for the three applications
The TOP 3 cases differ in their sensors but share the same layered structure. The value measurement layer is the one most often missing from publications.
Figure 5. Reference architecture for AI in a non-ferrous metals plant
Own work based on cases [1], [8], [14]–[19], [41], illustrative diagram. Layer 7 is explicitly described only in the EGA publication [1].
Poland and the EU
Where the industry stands
In 2025, AI was used by 14.3% of enterprises in divisions C24–C25 in the EU and 7.2% in Poland [28]. In production processes alone, the shares were 2.98% in the EU and 2.42% in Poland; Eurostat does not publish confidence intervals for these figures, so the difference should not be interpreted without them.
Figure 6. Share of enterprises (10+ persons employed) using at least one AI technology, 2024 and 2025
Source: Eurostat, datasets isoc_eb_ai (total) and isoc_eb_ain2 (by NACE section), updated 15.06.2026 [28]. Divisions C24–C25 jointly cover the manufacture of basic metals – including iron and steel – and of metal products, and are not representative of non-ferrous metals alone; Eurostat does not publish data for group 24.4.
Directions for 2027
What 2027 will change
Ten directions. Each states the type of basis: binding law, funding call, state policy objective, team forecast or vendor announcement.
Figure 7. 2026–2028 deadlines relevant to data and AI in non-ferrous metals plants
Sources [23]–[27], [29], [34], [38]. AI Act dates according to Regulation (EU) 2026/1744; Machinery Regulation – date after the corrigendum (OJ L 169, 04.07.2023). CRA vulnerability reporting obligations apply from 11.09.2026.
Binding law18.02.2027 · 31.12.2027
Battery passport and mass balance
The passport (Art. 77) is the responsibility of the economic operator placing the battery on the market; recyclers are subject to targets for recovery of materials – Co, Cu, Ni and Pb (90%) and Li (50%) – by 31.12.2027. From 2028 (18.08.2028 at the earliest), manufacturers of industrial batteries with a capacity greater than 2 kWh, electric vehicle batteries and SLI batteries must document the share of Co, Pb, Li and Ni recovered from waste (Art. 8) – the data come from recycling. On 11.09.2026 the Commission confirmed that it is not changing the targets [24].
To prepare: batch identifiers from feedstock to product, a mass balance model, an Asset Administration Shell interface (IEC 63278).
Binding law01.02.2027 · 30.09.2027
CBAM and the phase-out of free allocation
The CBAM declaration is submitted by the authorised CBAM declarant for imports; the mechanism covers aluminium, among others, but not copper, zinc or lead [25]. For EU producers, what matters is emissions data from suppliers of non-EU feedstock and the phase-out of EU ETS free allocation for CBAM sectors in 2026–2034 [37].
To prepare: energy and emissions measurement at installation and process level, an audit trail for emission-estimating models.
Binding law21.05.2027
Waste exports outside the OECD
From 21.05.2027, exports of non-hazardous waste to non-OECD countries are permitted only to countries on the Commission’s list; the draft list was open for consultation 18.09–16.10.2026 [38]. More scrap will remain in the EU – sorting to alloy level is gaining in importance.
To prepare: a forecast of scrap supply by fraction and quality, sorting capacity.
Funding call02.02.2027
Hybrid models and energy flexibility of electrolysis
TRIMET in Essen rebuilt 120 pots so that the potline’s power draw can be regulated by ±25% [35]; the critical quantity is the thickness of the frozen ledge on the side walls, which the model must estimate. The Horizon Europe call MAT-PROD-32 funds renewable energy and energy management in process industries [29].
To prepare: high-resolution time series covering 2–3 years; pot safety limits set by process engineers.
Team forecast2026–2027
AI sorting on industrial lines
Vendors are extending deep-learning platforms to further fractions, including copper–steel composites in steel scrap [10]. The advantage will go to the plant that measures purity and yield independently.
To prepare: reference samples, validation at the melt, a requirement in the purchase order to report metrics per fraction.
Vendor announcementfrom 2026
Supervised AI agents
Control system vendors are introducing AI agents for process data analysis [39]; in 2026 Horizon Europe funded AI in process industries, including agents (call closed). Principle: the agent proposes, a human approves actions that affect the process.
To prepare: controlled APIs to the historian, MES and CMMS; a register of AI applications; an owner for each agent.
Binding law03.04.2027 · 11.12.2027
OT cybersecurity as a prerequisite for AI
The amendment to the Polish National Cybersecurity System Act (KSC) (Dz.U. 2026 poz. 252, in force from 03.04.2026) gives essential and important entities 12 months to meet the obligations under Chapter 3 [34]; the CRA applies in full to manufacturers of devices and software from 11.12.2027 [26].
To prepare: segmentation according to IEC 62443, an inventory of devices and models, update clauses in purchase orders.
Vendor announcement20.01.2027 · 02.08.2028
Digital twin of the shop floor – pilot first
Factory twin platforms were announced in 2025 [40], with no reference installations in non-ferrous metals. AI as a safety component of a machine will be assessed under the Machinery Regulation (applicable from 20.01.2027); the AI Act obligations for such systems apply from 02.08.2028 [23][26].
To prepare: 3D models of halls, crane data; computing capacity of the PIAST-AI (Poznań) or Gaia (Kraków) AI factories.
State policy objectiveChina · 2027
China: AI as an element of industrial policy
MIIT’s “AI + manufacturing” plan (279号, 12.2025) envisages, by 2027, 1,000 industrial agents, 100 datasets and 500 typical scenarios, including industry models for non-ferrous metals; the aluminium plan (03.2025) requires new capacity to consume no more than 13,000 kWh/t (AC, for liquid aluminium) [36].
Binding lawfrom 02.02.2025 · 02.12.2027
AI literacy and the AI Act at the plant
Art. 4 of the AI Act requires measures to support the AI literacy of staff. The typical TOP 3 applications are not among the high-risk systems listed in Annex III; Annex I is relevant where AI is a safety component of a machine (from 02.08.2028) [23]. In 2025, 92.8% of Polish enterprises in divisions C24–C25 did not use AI [28].
To prepare: role-based training; a policy for using general-purpose models without transferring process data.
Limits of the conclusions
Limitations of the review and what would change the assessment
Limitations of the review
Publication bias: companies disclose successful deployments, while failed pilots remain unpublished; the absence of negative cases does not mean that none exist.
Results reported by companies and vendors have not been independently verified, except where indicated.
The review favours plants with well-developed communications; the WEF Global Lighthouse is an application-based programme.
Chinese sources are available mainly through news agencies and official lists, which reduces comparability.
The assessment was carried out without pre-registration of the protocol.
What would change the assessment
A processor publishes a vision or prediction result with a control group – TOP 3 would move up or be replaced.
Independent measurement of the purity of an AI cascade confirms a gain over LIBS alone – this would strengthen TOP 2.
Short-circuit detection in copper electrorefining is described as an AI method with a result – it would become a candidate for TOP 3.
Data on failed soft-sensor deployments emerge – this would weaken TOP 1.
Glossary
Terms
Soft sensor
a model that estimates a quantity measured infrequently or at high cost (e.g. bath temperature) from continuously measured quantities.
Anode spike vs anode effect
a spike is an anode protrusion approaching the metal (loss of current efficiency); an anode effect results from Al₂O₃ depletion (voltage jump, PFC emissions).
MPC
model predictive control: optimisation of setpoints over a time horizon subject to process constraints.
Hybrid model
a combination of process physics equations with machine learning; usually behaves better outside the range of the training data, provided that the physical part is correct.
LIBS
laser-induced breakdown spectroscopy: elemental composition of the surface layer of a piece (a few µm) in milliseconds – hence the sensitivity to coatings; Dynamic LIBS uses cleaning pulses.
XRT
dual-energy X-ray transmission: distinguishes materials by effective atomic number, independently of piece thickness.
Recall and precision
the share of real events that are detected, and the share of correct alarms among all alarms.
Edge inference
running the model on a computer at the line – with short, predictable latency and no dependence on network links or the cloud.
Reference group
units without the deployment, compared with units with the deployment over the same period – the basis for attributing the effect.
[6]Constellium (2024). Revolutionizing aluminium recycling: automotive scrap sorting by LIBS. constellium.com· company · N2
[7]EGA (15.01.2025). EGA designated Industry 4.0 Global Lighthouse by World Economic Forum. media.ega.ae· company · N2
[8]Novelis (2026). Sustainability Report 2025, p. 26. novelis.com· company · N2
[9]TOMRA (25.06.2025). TOMRA and Novelis webcast highlights the future of aluminum recycling. tomra.com· vendor · N3
[10]TOMRA (31.03.2025; 07.05.2026). GAINnext for aluminium; TOMRA launches next-generation AI platform and expands GAINnext ecosystem. tomra.com· vendor · N3
[11]Sortera Technologies (21.05.2026). Sortera Technologies doubles capacity. Company press release (PR Newswire). · company · N2
[12]Díaz-Romero D. J. et al. (2023). Resources, Conservation & Recycling 190, 106865. doi:10.1016/j.resconrec.2023.106865 doi.org· journal · N1
[13]Hydro (25.09.2024). Hydro and PADNOS start advanced aluminium scrap sorting operations in the U.S.; Dormagen and Nowa Sól plant web pages. hydro.com· company · N2
[16]MIIT (2024). List of “excellence-level” smart factory projects, batch I, item 229; Ningbo Jintian, annual report 2025, pp. 19–21. miit.gov.cn· authority and company · N2
[17]Southwire (2025). 2025 Sustainability Report, pp. 22, 25, 56. southwire.com· company · N2
[18]HES-SO Valais, Idiap, Constellium Valais (2021–2022). Project P3 – Press Pressure Prediction; SSRN preprint (2024), doi:10.2139/ssrn.4741197. idiap.ch· research institute · L
[30]MIIT (03.2025). List of typical cases of AI applications in new industrialisation; Keji Ribao, reprinted 11.06.2026: model “绿铝云慧鉴”. finance.sina.com.cn· authority and state media · N2†
[31]Reuters (26.08.2026). EGA restarts quarter of Al Taweelah smelter (reprinted by Kitco). kitco.com· news agency · context
[32]Lindthaler M., Filzwieser A. (2025). Digitalized state-of-the-art short-circuit detection (ESCD). 12th International Copper Conference, Springer. doi:10.1007/978-3-032-00102-3_23 doi.org· conference
[38]Regulation (EU) 2024/1157 on shipments of waste, Arts 39–41; European Commission, “Have your say” consultation 14861 (Ares(2026)8858290). eur-lex.europa.eu· legal act
[39]ABB (17.04.2026). Press release: agentic AI in Genix Datalyzer and My Measurement Assistant+. new.abb.com· vendor · N3
[40]NVIDIA (06.01.2025; 28.10.2025). Mega Omniverse Blueprint; Siemens as the first provider of digital twin software compliant with the blueprint. nvidianews.nvidia.com· vendor · N3
[41]Hindalco Industries (2026). Integrated Annual Report 2025-26, pp. 86–87. hindalco.com· company · N2
[42]Tongling Jingda Special Magnet Wire (2026). Annual report 2025, pp. 11–12; MIIT (2023), list of demonstration factories, item 52. static.cninfo.com.cn· company · N2
[43]Ningbo Powerway Alloy Material (2026). Annual report 2025, pp. 11–12. static.cninfo.com.cn· company · N2
About this study
Version, use of AI, interests, corrections
Version 0.9, as of 25.10.2026. Each case in the register carries the date of its last verification; a full review is carried out every 6 months, the next by 30.04.2027.
Use of AI tools in preparing the study
The source searches in six languages, the document summaries and the first draft of the text were produced with the help of AI tools – large language models operating as search and analysis agents.
Every figure and date was checked directly against the source document. The six fact audits, the technical audit and the scientific review were also conducted with the use of AI tools; discrepancies were resolved according to the primary source.
The scope, criteria K1–K5, P/N evidence ratings and final wording were adopted after editorial review.
No data on members of the Chamber and no non-public plant data were provided to AI tools – the study is based solely on the public sources listed in the bibliography.
The figures are illustrative diagrams prepared on the basis of the sources, not images generated by AI.
Errors, including incorrect attribution of a figure to a source, may be reported to the address given below; corrections are entered in the register with a new verification date.
The Chamber brings together companies, some of which are mentioned in this study or use the technologies described. The study was not funded by technology vendors.
We use cookies that are necessary for the website to work. With your consent we may also use analytics cookies to understand how the site is used. You can accept all, reject optional cookies, or set your preferences. More in our Privacy & Cookie Policy.