Supply Chain Management (SCM) provides an integrated, cross-organizational perspective on all business processes, encompassing key areas such as procurement, production, distribution, marketing, and controlling. These areas are crucial to the value chain, directly influencing efficiency and competitiveness. Optimizing them is essential for long-term business success. In this context, machine learning (ML) and deep learning (DL)-based artificial intelligence (AI) models offer advanced tools for enhancing these processes through data analysis and processing. Although advanced AI technologies and algorithms have become theoretically accessible to enterprises of all sizes through open-source software, the practical implementation of these sophisticated models remains a significant challenge.

There is a significant research gap in the area of practical applications of AI models in SCM that complicates a clear understanding of specific use cases. This paper provides an overview of current and potential AI techniques, focusing on experimental studies, case studies, and practical problem-solving approaches that can enhance SCM.

A Multivocal Literature Review was conducted, incorporating formal literature from databases such as ScienceDirect and Scopus, as well as grey literature from sources like Google. The search process followed an adapted approach based on Denyer and Tranfield (2009) and integrated elements of the PRISMA framework (Page et al., 2021) to ensure quality and transparency. Identification of grey literature followed the method pruposed by Garousi et al. (2018). The Google search was restricted to PDF files, and the top 200 search results were reviewed. In total, 1,454 reports were examined, with 121 included in the qualitative analysis.

The results reveal that the most frequently employed AI technique is from the field of deep learning, specifically artificial neural networks in various forms (FFNN, CNN, RNN), which were utilized in 71% of the reviewed literature and applied across all investigated SCM domains. These deep learning approaches significantly contribute to quality management, pattern recognition, reliability analysis, demand forecasting, fault diagnosis, and improvements in risk management and resilience.

This study provides important insights into the challenges and motivation in AI adoption. A central challenge, particularly within SCM environments, is the necessary collaboration among multiple stakeholders and the management of heterogeneous data sources. Additional challenges include trust in AI outcomes, concerns about data security, issues in change management, and the lack of suitable frameworks for AI adoption and governance. The motivations for AI implementation are diverse, with performance optimization being a primary focus. Additionally, ecological motivations, such as promoting sustainability in production, and enhancing resilience are also noted.

This paper provides a novel perspective on SCM and emphasizes the critical role of AI.

Global manufacturing companies need to incorporate environmental, social, and economic criteria to enhance their competitiveness and sustainability. These metrics are becoming increasingly important. The movement of commodities, the provision of services, and logistical processes are all areas in which supply chain operations have a substantial impact on the environment. Important strategic decisions, particularly in carrier and mode of transportation selection, significantly impact the efficiency of logistics. Beneficial Cargo Owners (BCOs), or the companies whose goods are being delivered, must carefully examine several variables when choosing a carrier, including cargo security, equipment availability, transit times, and freight costs. These decisions are crucial because they have an immediate impact on customer satisfaction and operational effectiveness. To track cargo and manage inventory, BCOs mostly rely on real-time data. This allows them to guarantee accurate and timely deliveries, which improves the overall customer experience and satisfaction levels. In response to global imperatives and stakeholder expectations, manufacturing giants are increasingly acting upon sustainability mandates reflected in their reports. These actions include managing shipment volumes, adopting alternative energy sources, and actively pursuing strategies to mitigate emissions across their supply chains. The quest for more effective strategies at the strategic level is driving BCOs to embrace innovative approaches such as the implementation of multimodal freight transportation models. Through technology integration, service quality, and environmental impact factors, carrier selection criteria are optimized in this manner. The goal of the approach is to help BCOs achieve more sustainability and efficiency in their supply chain operations by emphasizing the relationship between technology, service quality, and environmental responsibility. They can minimize their environmental impact while optimizing operations through strategic decisions in carrier selection and logistics. Companies that adopt such models are positioned strategically to satisfy changing customer expectations for environmentally friendly operations and to comply with industry standards and regulatory regulations. Examining these important parameters in detail, this study suggests an all-encompassing, sustainable multimodal freight transportation optimization model. Two integrated decision support mechanisms are included in the model: a route generation mechanism that generates different routes based on the limitations of the BCO and a multi-criteria decision-making system that chooses the most appropriate carrier. At this stage, a mixed-integer programming model is created to produce multimodal transportation routes after a multi-criteria decision-making analysis is completed on the established carrier selection criteria. Based on the outcomes of the multi-criteria decision-making study, the model's parameters and decision variables are assigned. As a future direction of the study, an application phase using case study data will be carried out to validate the generic model that is intended to be fitted to operating BCOs in various sectors. As a result, this study positions itself to provide a way forward for improved operational effectiveness, lower expenses, and more customer satisfaction in the ever-changing world of international trade.

During outbound rail transportation from seaport terminals, containers are secured to the rail wagons using foldable pins to prevent slipping. In seaport terminals, depending on the container train loading scheme, the container pins must be flipped up or down before commencing container loading operations. The research project "Pin-Handling-mR" aims to automate this process using a mobile robot. The automation is expected to enhance the safety of terminal staff and reduce operational costs for terminal operators. A key challenge to enable this automation is the accurate recognition and pixel wise segmentation of container pins, that is essential for precise robotic manipulation. Deep learning frameworks can be leveraged to accurately detect, classify and segment container pins. However, such frameworks require high-quality annotated data for training, and to the best of author’s knowledge, no publicly sourced container-pin dataset exists to train and test a deep learning image segmentation network for real time container pin segmentation.

This study focuses on the evaluation of the state-of-the-art deep learning based models to detect and segment the pins with better accuracy. It proposes the creation of a Red-Green-Blue-Depth (RGB-D) container pin dataset, featuring two types of container pins, and the development of a deep learning-based object segmentation pipeline. The dataset is generated using images from two sources – first, a demonstrator table equipped with container pins and an RGB-D camera and second, images captured from the Container Terminal Tollerort (CTT) in Hamburg, Germany. Two state-of-the-art deep learning networks, namely Mask-RCNN and CMX, are trained by fusion of RGB and depth data from dataset and fine-tuned using a transfer learning approach. The performance of these networks is compared using evaluation metrics for both RGB and RGB-D data. Validation and final tests are performed through segmentation tests on images collected from CTT. The final trained networks achieve accurate classification and generate precise pixel-level category masks for both RGB and depth modalities.

The results demonstrate that the best-performing network has achieved an Average Precision (AP) of 92.5%, an Average Recall (AR) of 94%, and category-wise APs of 92.3%, 88.29%, and 97.1%. Such successful inference confirms the viability of deep learning-based frameworks for container-pin segmentation tasks.