Risk-sensitive collaborative parameter tuning via calibrated deep surrogates for rare-earth electrolysis energy efficiency

Improving energy efficiency in rare-earth molten salt electrolysis is challenging due to strong process nonlinearity, limited evaluation budgets, and strict operational constraints. Existing data-driven parameter tuning methods often rely on deterministic surrogate models and pointwise optimization, which ignore predictive uncertainty and may result in aggressive decisions and increased risk of constraint violations. To address these issues, this paper proposes a risk-sensitive collaborative parameter tuning framework based on calibrated deep surrogate modeling and uncertainty-aware constrained Bayesian optimization. A deep ensemble surrogate is employed to predict energy-efficiency indicators and quantify uncertainty, while a calibration procedure is introduced to improve the statistical reliability of predictive uncertainty. Based on the calibrated surrogate, parameter tuning is formulated as a risk-sensitive constrained optimization problem and solved efficiently using an uncertainty-aware acquisition strategy. Numerical experiments show that the proposed framework achieves faster convergence, higher energy-efficiency improvement, and lower constraint violation probability than uncertainty-agnostic alternatives, demonstrating its effectiveness for safe and efficient electrolysis operation.