Answers are grounded in this paper and cite their sources.
The aluminium industry is among the most carbon-intensive metals sectors, and recycling which saves roughly 95% of the energy used in primary production is the principal route to its decarbonisation. Operational planning for recycled aluminium supply chains, however, remains tied to models that assume stable demand, fixed carbon costs, and a one-time analytical solution, ignoring the volatility encountered in practice. This paper develops a more realistic framework for a three-echelon recycled aluminium product supply chain by combining analytical optimisation and reinforcement learning to derive both interpretable benchmarks and adaptive decisions. Three modelling extensions are introduced. First, demand responds simultaneously to price, displayed inventory, and green-technology investment. Second, carbon allowance prices are modelled using a stochastic Ornstein–Uhlenbeck process consistent with cap-and-trade market behaviour. Third, a reinforcement-learning agent learns order quantity, cycle time, and selling price based on the realised market state. Solar-powered logistics and carbon capture and storage are retained from prior work. The deterministic optimum is computed analytically and verified through Hessian analysis, yielding ( 𝑄 ★ , 𝑇 ★ , 𝑃 ★ ) = ( 1 3 0 . 0 6 , 0 . 7 6 3 2 , 1 0 0 . 4 8 ) with an integrated profit of 𝐼 𝑇 𝑃 ★ = 9 , 7 4 8 . 1 7 , which the multi-driver demand raises by 3.7% over a single-driver replication. In the stochastic environment, evaluated over five independent training runs and 100 held-out episodes under common random numbers, the learned policy attains a mean per-period profit of 10, 287.49 against 10, 069.10 for a fixed-decision baseline, a paired improvement of 2.17% ± 0.65 percentage points ( 𝑝 < 1 0 − 8 ). A state-wise re-optimisation benchmark that observes the state exactly attains 4.00%, so the learned policy captures approximately half of the flexibility value available while requiring no model at run time. The improvement is not uniform across training runs or operating regimes—retraining and re-evaluating the pipeline at thirteen alternative demand and quality settings shows the learned policy losing significantly to the fixed policy in three of them despite positive flexibility value being available in all thirteen and, contrary to what might be expected, adaptation raises rather than lowers profit volatility, because expected-profit maximisation expands output in favourable states. Sensitivity analysis identifies consumer responsiveness to green credentials and product quality as the dominant profit drivers, subject to the caveat that the demand elasticities are calibrated rather than estimated. The framework pairs a certified analytical benchmark with adaptive decision-making under uncertain carbon prices and weather, and quantifies both the value and the limits of adaptation for green branding, working-capital management and policy design.
No takes yet. Share an insight, caveat, or question.
Bhadoriya et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: