Answers are grounded in this paper and cite their sources.
Digital twins need data that are physically consistent, temporally realistic and honest about uncertainty. For end-of-life permanent-magnet recycling, central to rare-earth supply security yet poorly characterised, they are absent. Generated data can substitute, but resemblance, utility and privacy criteria are defined against the missing data themselves, so a synthetic foundation cannot be certified. This study builds one for an eleven-process neodymium-iron-boron and ferrite recycling chain. The generator combines autoregressive noise, dual seasonality, a wear trend, noisy conversion ratios and two-pass mass-balance conservation over 500 replicates. Seven structural fits recover the configured ratios to within 0.11% (R² 0.979-0.999), and 85.4% of an 82-model screen satisfy five diagnostic criteria (best-per-process R² 0.24-0.98). Four machine-learning families all degrade mean R² (0.000-0.041), as a linear generating process predicts; this linearity-recovery test is proposed as a within-synthetic criterion of internal consistency, not of external validity. Chain demand is 442.1 kWh per hour and energy intensity 9.58 kWh per kilogram of magnet product against 1.19 for ferrite; no saving against primary neodymium is claimed, the pellet's composition being unresolved. Relative dispersion rises from 3.46% to a 4.62-4.69% plateau, so conservation bounds variance propagation. Its cost follows in closed form: renormalisation attenuates each stage's injected dispersion by a factor set by its conversion ratios, producing the near-unit coefficients of determination and erasing the specified seasonal and wear-trend structure. With near-uniform dispersion, a binary-knapsack reallocation of the control budget funds high-throughput upstream stages, not the energy-dominant Furnace: the uncertainty and energy layers order the chain oppositely.
No takes yet. Share an insight, caveat, or question.
Sifakis et al. (2026) studied this question.