UC Berkeley IEOR-led Team Takes Second Place for Predictive Model Aiming to Improve Rail Safety

UC Berkeley IEOR PhD student Alberto Gennaro and his team are applying data-driven modeling to a problem that directly affects the safety and reliability of railroads: predicting when train wheels are likely to fail. Their project earned second place in the 2025 INFORMS Railway Applications Section Problem Solving Competition, which invites researchers and practioners to create predictive maintenance tools for rail systems. Using wheel profile data, mileage, railcar attributes and wayside detector inputs such as WILD and truck hunting measurements, teams work to forecast wheel failures and improve rail safety through advanced analytics and machine learning.

Gennaro collaborated with a multidisciplinary team that included Daniele Gioia, an applied scientist with Zalando and the German Aerospace Center (DLR); Edoardo Fadda, an assistant professor at Politecnico di Torino; and Jacopo Bonari, an engineer at DLR. Together, they combined expertise in operations research, mechanical engineering, data science, and mathematics to build a system capable of anticipating wheel failures in real-world rail environments

Building on guidelines from the Federal Railroad Administration and the Association of American Railroads, the researchers designed a nested classifier that first assesses the likelihood of wheel failure and then categorizes the failure mode. Their work relied on Bayesian methods and required substantial exploratory analysis, as the dataset contained noisy readings and missing values. The team devoted considerable effort to sampling strategies, gap filling and feature construction to improve the model’s reliability.

Gennaro said the project’s progress depended on pairing rigorous statistical modeling with disciplined data engineering to make use of imperfect real-world data. He credits his UC Berkeley IEOR training with informing the model’s structure and supporting effective collaboration across disciplines and time zones.