Disruption-resilient lot-sizing and scheduling using multimodal deep reinforcement learning and distributionally robust evaluation

Modern manufacturing systems face increasing pressure to remain cost-efficient, responsive, and resilient amid operational uncertainty and disruptive events. While traditional scheduling methods offer limited adaptability, existing AI-based solutions often overlook robustness and risk sensitivity in complex production settings. This study addresses this gap by developing a multimodal deep reinforcement learning (MDRL) framework that learns data-driven scheduling policies under routine stochastic variability and evaluates them through distributionally robust and tail-risk-based disruption stress tests. The proposed approach leverages a transformer-based algorithm enhanced with transformer encoders and cross-modal attention to processing diverse data sources. Robustness is assessed through scenario-based disruption stress testing and CVaR/DRO-based robustness scoring for post-training checkpoint selection. Experimental results reveal that the MDRL agent consistently favors mid-sized lot sizes (peaking around 58 units) and reduces setup activations by approximately 30%, promoting stable and efficient operations. Analyses for robustness show that while increasing the CVaR threshold from 0.5 to 0.99 raises worst-case cost by over 50%, it leads to a 26% drop in service level, highlighting critical trade-offs between risk aversion and responsiveness. These findings highlight the potential of MDRL to optimize operational performance while supporting systematic robustness evaluation in real-time production environments. The framework offers a scalable and intelligent decision-making tool for next-generation smart factories, with implications for operational efficiency, sustainability, and supply chain resilience.