Can AI Cut Scrap and Waste?
Every scrapped part has already used its material, energy and machine time. Machine learning promises to find better process settings in fewer trial runs and to warn when production drifts. The factory results we found are few, the strongest co-written by the manufacturer, and they hold only inside the window where each model was tested.
On this page
Better: Note and confidence
Low to moderate confidenceSteadier settings inside a validated windowOne field study co-written by Beko staff reported less scrap on two moulded parts over six months of use, without stating the baseline. In one lab study, models predicted part properties better than the settings needed to reach them.Faster: Note and confidence
Low confidenceFewer trials to a recipe, slow to set upIn Lam Research's simulated etch benchmark, an expert followed by an algorithm reached target for about half the expert's cost. A lab fibre process found its best settings within 14 of 20 iterations but missed one of three targets. In the Beko-authored study, defining the reliable zone first took simulation and designed experiments.Cheaper: Note and confidence
Low confidenceSaves material where the model is validated and engineers leadIn Lam Research's simulated benchmark, an expert followed by an algorithm reached target for just under half the expert's cost, but algorithms alone beat the expert in 13 of 300 attempts. We found no independent multi-plant cost study.Safer: Note and confidence
Governance-gatedOnly with validation and change controlA model that changes process settings changes how a product is made. In our reading, removing a test on the strength of a model needs documented credibility evidence, a named owner and revalidation when inputs change.The problem, in numbers
Australian manufacturing generated an estimated 12.8 million tonnes of solid waste in 2018-19, 16.9% of the national total and narrowly ahead of construction, after a 14% rise in two years.Source 1 It produced 116 tonnes of waste for every million dollars of value added, against 87 for construction, and spent $1.2 billion on waste services.Source 1 These are experimental estimates, now several years old.Source 1
Food has its own, more detailed estimate. An updated national food waste baseline, prepared for Food Innovation Australia in 2021 and based mainly on 2018-19 data, estimates that food manufacturing produced 1.276 million tonnes of food waste out of a national 7.676 million, while a further 9.854 million tonnes was recovered, most of it as animal feed. Only 79,000 tonnes of the manufacturing waste went to landfill.Source 2 That is a modelled estimate, 17% lower for manufacturing than the 2019 baseline estimate, a comparison of estimates rather than a measured fall.Source 2 Because so little goes to landfill, in our reading the saving on offer is less about disposal fees than about not making the loss at all.
Our reviews of quality inspection and line throughput cover catching defects and lifting output. This one asks a narrower question: can AI tune process settings so that fewer bad parts are made in the first place?
What improves, and why
In our reading, some scrap starts as a setting that no longer suits the material or the tool. Engineers often find workable settings by trial runs, and every trial consumes material. The case for machine learning is that a model linking settings to outcomes can choose the next trial more efficiently, and later warn when production drifts out of the window where good parts are made.
Trials are expensive. In semiconductor etch development, researchers at the equipment maker Lam Research put the cost at more than a thousand dollars per experiment for the wafer, tool time and electron microscopy, and say engineers typically test only on the order of a hundred recipes.Source 5 At Deakin University, a Bayesian optimisation method tuned a lab microfluidic process for making short polymer fibres; across nine runs capped at 20 iterations, the best result in each run came by iteration 14 or earlier.Source 6
The model does not remove the trial runs. It decides which ones are worth paying for, and it can only be trusted inside the range where it has been tested.
What the evidence actually says
One measured factory result. The strongest example we found is a peer-reviewed study by researchers from Koç University and the appliance maker Beko, so the manufacturer is partly reporting on itself. They fitted cavity pressure sensors to moulds, mapped a reliable zone for good parts, and trained a model to suggest which machine setting to change when production drifted out of it. Over six months of use at Beko and its suppliers, the authors report that the scrap rate on a dishwasher cover fell by 800 parts per million, scrap on a refrigerator part fell by 52%, and start-up scrap fell by 49%.Source 4 The authors chose these parts because their scrap rates were high, and the paper does not state the baseline scrap rates, how the reductions were calculated or any comparison line.Source 4
A controlled benchmark. Lam Research's study in Nature built a simulated etch process so that people and algorithms could compete on equal costs of $1,000 per recipe and $1,000 per batch. The best human participant, a senior engineer, reached the target for $105,000. Algorithms working alone beat that in only 13 of 300 attempts. When the expert went first and then handed over to an algorithm, the median cost fell to $52,000, just under half.Source 5 The process was virtual, not real wafers, and the authors work for an equipment supplier.Source 5
Single-plant and lab studies. At a Thai stainless-steel wire mesh plant, a model that matched settings to each incoming batch of raw material was tested in real production, with process yield reported at an average of 91.3%, though the abstract gives no baseline to compare against.Source 7 In a US university lab moulding recycled polypropylene, models predicted three of four part properties within 10%, but their predictions of the settings needed carried errors of 3% to 25%.Source 7
A peer-reviewed review of machine learning studies for production quality and process improvement from 2008 to 2018 found, according to its abstract, hardly any correlation between the data used, its amount, the algorithms and the production problem.Source 3 In our reading, results do not transfer by default; each line needs its own evidence.
Where it falls short
Set-up comes first. Beko's method needed simulation and designed experiments to define the reliable zone before the model could advise, which the authors call one of its biggest limitations.Source 4
Only valid where tested. The Deakin team say their results hold only within the experimental states they tested, and the diameter in one of three fibre targets was never reached.Source 6
Data gaps cap the gain. The wire mesh study found performance limited by the quality and completeness of historical data, and some beam lots still had defects.Source 7
Unstable runs. In the recycled polypropylene lab, the adaptive control was unable to stabilise one run, which the authors link to that run's process settings.Source 7
Where governance decides the outcome
A model that changes process settings is changing how a product is made. In our reading, where that product carries a safety or regulatory requirement, a setting change suggested by a model needs the same change control, sign-off and records as one made by hand. The value of formal validation shows in a peer-reviewed engineering study of a component manufacturing process: once the team had established the credibility of a model that combined machine learning with a physics-based adjustment, it had grounds to drop a costly intermediate testing step used to set a key control variable, according to its abstract.Source 8 In our reading, documented evidence can justify removing a check; a promising accuracy figure cannot. Record the window the model was validated in, keep a named person accountable for accepting its suggestions, and revalidate when the material, tool or machine changes.
How to think about adoption
Start with the parts whose scrap you already measure and cost. Instrument the process first; the Beko results depended on pressure sensors fitted inside the moulds.Source 4 Budget material for designed trials to establish where the model works. Pair the model with experienced process engineers: in Lam's simulated benchmark, the algorithms succeeded only when an expert first guided them towards a promising range.Source 5 For food manufacturers, the feasibility study's scenario modelling assumes that improving manufacturing processes and technologies, not AI specifically, could save 147,438 tonnes in 2030 under an industry-led scenario, which is a projection, not a measured result.Source 2 Measure scrap per unit before and after, on a comparison line where you can, and treat any supplier's percentage as a hypothesis to test.
BFCS.ai does not fabricate figures. The national waste figures are statistical estimates and the food waste figures are modelled. The factory and benchmark results here come from studies written wholly or partly by the companies involved, and the rest are lab or single-plant studies. We found no independent multi-plant study that measured lower scrap costs after AI process tuning, and we say so where it matters.
Sources
- 1
Australian Bureau of Statistics, "Waste Account, Australia, Experimental Estimates," reference period 2018-19, released 6 November 2020. abs.gov.au
Back to textWhat we took from it
Australia generated 75.8 million tonnes of solid waste in 2018-19; sectors generating the most were manufacturing 12.8 million tonnes (16.9%) and construction 12.7 million tonnes (16.8%); manufacturing waste increased by 14% since 2016-17; manufacturing spent $1.2 billion on waste services; waste intensity was 116 tonnes per $m of gross value added for manufacturing and 87 for construction.
- 2
Food Innovation Australia Ltd (FIAL), "The National Food Waste Strategy Feasibility Study: Final Report," 2021 (delivered by an Edge Environment consortium with Lifecycle Strategies, WRAP and 3 Keel; report pages dated 30 August 2021). openaccess.city.ac.uk
Back to textWhat we took from it
Read from a third-party copy in the City, University of London open-access repository, not from FIAL's own site. The baseline is primarily based on 2018-19 data. Table 3 (updated national food waste baseline, thousand tonnes): manufacturing total waste 1,276 of a national 7,676, of which 79 to landfill; manufacturing total recovery 9,854, including 6,253 to animal feed; material recovered as non-food product is excluded from waste. Section 4.2.3: primary and manufacturing waste were 26% and 17% lower in the 2021 estimate than in the 2019 baseline. Table 8 (intervention long list and assumed abatement figures as tonnes saved in 2030): "Improve manufacturing processes and technologies" 147,438 tonnes under the industry-led scenario; the intervention is not specific to AI, and the report says high-level scenario analysis "often relies on significant assumptions." These are modelled estimates and projections, not measurements.
- 3
Weichert, D., Link, P., Stoll, A., Rüping, S., Ihlenfeldt, S. & Wrobel, S., "A review of machine learning for the optimization of production processes," The International Journal of Advanced Manufacturing Technology 104, 2019, pp. 1889-1902 (peer-reviewed). doi.org
Back to textWhat we took from it
Abstract only (full text is subscription): the review covers literature from 2008 to 2018 on machine learning and optimisation for product quality or process improvement and finds "hardly any correlation between the used data, the amount of data, the machine learning algorithms, the used optimizers, and the respective problem from the production."
- 4
Arslan, E., Konuşkan, Y. & Lazoglu, I., "AI-driven cognition for advanced injection molding and industrial implementation," The International Journal of Advanced Manufacturing Technology 138, 2025, pp. 2043-2064 (peer-reviewed, open access). link.springer.com
Back to textWhat we took from it
Authors from Koç University and Production Technologies R&D, Beko Corporate. Cavity pressure sensors in the mould and a convolutional neural network suggest machine-parameter changes when production leaves a reliable zone; 98% accuracy for the baseline model and an average of 95% across the two industrial applications. Parts with high scrap rates were prioritised for field tests. Used at Beko Corporate and its suppliers for six months: scrap rate on a dishwasher bottom cover fell by 800 parts per million; in the paragraph on a refrigerator inner cover, scrap in part production fell by 52%, and scrap during production start-ups fell by 49% without re-adjusting machine parameters. The paper does not state baseline scrap rates or how the reductions were calculated, and describes no comparison line. The authors state that one of the biggest limitations is that both CAE (simulation) and DoE (designed experiment) studies must be conducted to determine the reliable zone.
- 5
Kanarik, K. J., Osowiecki, W. T., Lu, Y. et al., "Human-machine collaboration for improving semiconductor process development," Nature 616, 2023, pp. 707-711 (peer-reviewed, open access). pmc.ncbi.nlm.nih.gov
because the nature.com page redirected to a cookie sign-in. nature.com
Back to textWhat we took from it
All authors are from Lam Research Corporation, a semiconductor equipment supplier. Read via the PubMed Central copy (PMC10132970),
Process data on silicon wafers costs more than a thousand dollars per experiment for the wafer, plasma equipment operation and electron microscopy; engineers typically test only on the order of a hundred recipes. In a virtual plasma etch process game, costs were set at $1,000 per recipe and $1,000 per batch; the expert human benchmark (the winning human participant) was $105,000; algorithms alone beat the expert in 13 of 300 attempts; a human first, computer last strategy with the best algorithm reached a median cost-to-target of $52,000, "just under half the cost required by the expert alone."
- 6
Li, C., Rubín de Celis Leal, D., Rana, S., Gupta, S., Sutti, A., Greenhill, S., Slezak, T., Height, M. & Venkatesh, S., "Rapid Bayesian optimisation for synthesis of short polymer fiber materials," Scientific Reports 7:5683, 2017 (peer-reviewed, open access). pmc.ncbi.nlm.nih.gov
Back to textWhat we took from it
Deakin University with a co-author from HeiQ Australia; a microfluidic device run with syringe pumps. Nine runs of 20 iterations each across three targets; best results were found between iteration 2 and 14; for the third target the diameter deviation stayed at 104% and the diameter target was not achieved (the length target was); results "are only valid within the set of tested experiment states."
- 7
Wongwan, K. & Laosiritaworn, W., "Data-driven quality improvement in woven wire mesh production using machine learning algorithms," Journal of Industrial Engineering and Management 18(3), 2025 (double-blind peer review with at least two external reviewers, per the journal's editorial policy; open access). doi.org
Krantz, J., Licata, J., Raju, M. A., Gao, P., Ma, R. & Masato, D., "Machine Learning-Based Process Control for Injection Molding of Recycled Polypropylene," Polymers 17(7):940, 2025 (peer-reviewed, open access), University of Massachusetts Lowell; one 130-ton machine and a spiral-flow mould. pmc.ncbi.nlm.nih.gov
Back to textWhat we took from it
Abstract only: case study at a stainless-steel woven wire mesh plant in Thailand; a gradient boosted tree model identifies production settings to suit new incoming batches based on raw-material inspection data, and the predicted machine parameters were tested in actual production; process yield increased "to an average of 91.3%" (no baseline is given in the abstract); performance is limited by the quality and completeness of historical data, and some beam lots still had defects.
Yield stress, ultimate elongation and part weight were predicted within 5% (regression) or 10% (neural network), modulus less reliably; predictions of processing inputs had errors ranging from 3% to 25% depending on model and response; in one run the adaptive process control "was unable to stabilize," which the authors link to the longer trigger time in that run; the authors suggest additional input data may be necessary.
- 8
Banyay, G. A., Worrell, C. L., Sidener, S. E. & Kaizer, J. S., "Credibility Assessment of Machine Learning in a Manufacturing Process Application," Journal of Verification, Validation and Uncertainty Quantification 6(3):031007, September 2021 (ASME, peer-reviewed), doi 10.1115/1.4051717. pure.psu.edu
Back to textWhat we took from it
Abstract only, read on the Penn State research portal record: a framework for establishing the credibility of a machine learning model that predicts a key process control variable setting to maximise product quality in a component manufacturing application; the model coupled a data-based model with a physics-based adjustment encoding subject matter expertise; establishing credibility "provided the basis for eliminating a costly intermediate testing process."