Complex products such as aerospace components and semiconductor wafers face severe challenges including high demand uncertainty, excessive inventory costs, and weak real-time responsiveness in multi-stage production. Most existing AI-driven inventory models separate forecasting and control, limiting their adaptability to dynamic manufacturing environments. To address this gap, this study develops an integrated AI framework combining Long Short-Term Memory (LSTM) networks and Q-learning for intelligent decision-making and inventory optimization across the full lifecycle of complex product production. The LSTM module accurately captures long-term temporal dependencies in multi-source time-series data, achieving high prediction accuracy with significantly reduced error compared with traditional ARIMA. The Q-learning module enables dynamic adjustment of production and inventory via real-time interaction with the production environment. Experimental results based on real industrial data show that the proposed framework reduces total cost by 15.7% relative to conventional MRP systems and 8.3% compared with advanced MIP-DRL models, while achieving 0% stock-out rate and 4.2 annual inventory turnovers. It also exhibits strong robustness under different demand volatility levels. This study enriches the theory of AI-driven multi-stage optimization and provides a scalable paradigm for high-value manufacturing to enhance operational efficiency and cost-effectiveness.
Jiang et al. (Thu,) studied this question.