Machine Learning for Multimodal Freight Chain Modelling

Abstract

Freight transport plays a critical role in supporting global trade, with multimodal transport systems modelling gaining importance due to their potential to optimize efficiency and sustainability. Accurately modeling these multimodal freight chains is essential for infrastructure planning and policy-making. Yet, it remains a persistent challenge due to fragmented datasets, limited granularity, and the absence of observed multimodal chain-level data. Traditional modeling approaches, particularly heuristic-based methods, often struggle to incorporate real-world operational constraints such as port selection logic and cargo handling requirements. Moreover, these models are typically inflexible and computationally intensive.

This research seeks to develop an adaptive multimodal freight chain model that addresses these limitations. Specifically, it introduces a practical path construction framework that integrates port selection based on geographic and functional suitability, aligning cargo handling requirements with port capabilities during the construction of mode chains. This research also tries to address a gap in the literature by applying machine learning to the estimation of multimodal freight flows, a domain traditionally dominated by heuristic and optimization-based methods. To estimate freight demand distribution across the generated chains, this study explores the use of machine learning, particularly the Expectation-Maximization (EM) algorithm, to leverage the abundant but often unstructured transport data available. The EM model enables demand share prediction without relying on labeled training data, reducing the calibration burden and enhancing model responsiveness to observed transport flows.

The proposed modeling framework is applied to a case study based on the NEAC Mode Chain Builder system for inter-country freight movements between the Netherlands and Belgium, two countries with high multimodal connectivity and the largest ports in Europe. The results demonstrate the model’s ability to generate valid mode chain alternatives and to significantly reduce deviations between predicted and observed freight flows, particularly for sea and rail segments. While some deviation increases occur in other segments, these are outweighed by the overall improvement in prediction accuracy. The EM model also shows stable convergence behavior, confirming its potential under data-limited conditions. However, residual deviations suggest that external factors, such as data incompleteness or behavioral uncertainties, still limit full accuracy.

This study highlights the potential of combining graph search algorithms with unsupervised learning to enhance multimodal freight chain modeling, especially under data-constrained conditions. It contributes both a methodological and practical solution for building data-driven multimodal freight transport models that better reflect operational realities and observed empirical data that can be used to improve the freight transport planning and decision-making process.