This paper presents a novel framework for integrating deep learning (DL) into transport logistics system dynamics (SD) modeling. To address the insufficient explanatory power and low causal reliability of existing DL-based models, we propose a hybrid approach that combines concept-based interpretability, mechanistic interpretability, and causal machine learning techniques with DL. This approach builds a neural network model using semantically explicit and actionable variables, preserving the causal basis and transparency of existing SD models. We validate this approach by applying it to a real-world case study from the EU AutoMoTIF project (data-driven decision support, automation, and optimization of a multimodal logistics terminal), demonstrating the role of neural symbolic methods in bridging the gap between black-box predictive models and critical decision support in the complex dynamic environment of industrial IoT-based cyber-physical systems.
