Machine learning for risk management in construction projects

Risks and uncertainties are inherent in construction projects, presenting challenges to their successful completion and outcomes. Conventional risk management practices often rely on manual, time-consuming, subjective, and experiential processes, leading to superficial and ineffective risk identification and assessment, and hindering the knowledge transfer to the next projects. To overcome these limitations, this research aims to explore the potential of Artificial Intelligence (AI) and Machine Learning (ML) algorithms to automate and optimize risk management processes in the construction industry. The objectives of this research are as follows: identifying the shortcomings of traditional risk management practices, defining and implementing suitable ML algorithms, addressing data scarcity and uncertainty issues in construction companies, representing interdependencies between project variables and risks using probabilistic graphical models, automating risk identification and assessment processes, continually assessing and improving risk management performance, and addressing practical implementation requirements, challenges, ethics, biases, and potential harms. To achieve these objectives, three distinct ML-based models were developed and applied to two case studies. The first case study involved analyzing the project portfolio of Jacobs Italia SPA, the industry partner of this research. The second case study utilized a comprehensive database comprising over 130,000 records of school buildings in New York City. The models employed were a probabilistic Bayesian Network model based on both subjective expert opinions and objective project data, a Fuzzy Logic model based solely on subjective expert data, and a deterministic ML model based solely on previous project data. The results from each model were recorded and analyzed to assess the influence of database size on the performance of the ML models, as well as the role of uncertainty in achieving more accurate and realistic risk predictions. While deterministic ML models with backbox structures performed better in the bigger database, Bayesian Networks demonstrated the most favorable performance in limited databases, suggesting that the integration of subjective data elicitation with objective data is an effective approach to compensate for data scarcity. Another proposed solution to address data scarcity was the use of data augmentation and synthetic data generation through Generative Adversarial Networks (GANs), which proved highly effective. The findings indicate that ML models, particularly probabilistic ones like Bayesian Networks, can significantly enhance risk management processes across various project knowledge areas. These models facilitate the identification and analysis of complex interrelationships and causal inferences between project variables, providing accurate estimates of potential risks. As a result, project managers are empowered to take proactive measures and make informed decisions to mitigate risks, leading to successful, safe, on-budget, and on-time project delivery. Although limitations such as data scarcity and a lack of benchmark studies exist, the proposed ML-based risk management framework offers several contributions to the risk management body of knowledge, providing practical implication solutions and proper ML model choices based on the requirements and resource availability of each company. These include identifying key risk features, providing a comparative analysis of ML algorithms, addressing implementation challenges, and promoting the adoption of AI in the construction industry.