Regime-Aware Mineral Price Prediction via Hybrid Quantum–Classical Deep Learning with Mid-Layer Attention Fusion

Abstract Critical mineral prices are volatile, regime-shifting, and highly reactive to policy interventions, supply stress, and speculative dynamics. In materials and manufacturing supply chains, such volatility affects procurement timing, inventory buffers, and cost-risk exposure; therefore, one-step-ahead prediction is not merely extrapolation; it requires inferring the active latent regime and adapting rapidly to abrupt transitions. This paper proposes an attention-guided quantum–classical hybrid architecture for next-step prediction of such nonstationary industrial time series. The model consists of two coordinated branches. A classical bidirectional long short-term memory predictor models medium-horizon temporal structure from historical windows. In parallel, a quantum-gated bidirectional long short-term memory model replaces classical gate computations with a parameterized quantum circuit, mapping the current multivariate input and latent state into a quantum feature space to produce context-dependent gate activations designed to be sensitive to structural breaks. The two representations are fused through a mid-layer attention mechanism that (i) adaptively weights the classical and quantum branches at each time step based on inferred reliability and (ii) applies temporal attention to emphasize the most informative subsequences of recent history for the final prediction. To strengthen regime awareness, the revised framework incorporates multivariate endogenous inputs together with explicit regime-conditioned features derived from a training-only clustering procedure. Experiments on daily lithium carbonate prices show that the proposed hybrid improves one-step prediction accuracy and calibration relative to either branch alone, with consistent improvements in accuracy and calibration across both stable and volatile market regimes. Additional validation on gallium exhibits the same qualitative pattern, supporting the transferability of the regime-aware fusion strategy beyond a single mineral series. Overall, the results suggest a practical design pattern for critical mineral prediction in which the quantum branch acts as a complementary nonlinear representation module whose contribution is mediated by the attention-fusion mechanism, while classical deep learning provides stable temporal modeling for decision-relevant forecasting in volatile critical-mineral markets. An attention-weight analysis indicates that the fusion mechanism does not selectively amplify the quantum branch during volatile or transition periods; the observed improvement is, therefore, best understood as a fusion-level effect rather than as standalone superiority of the quantum branch.