Prediction of Bridge Inventory Characteristics using Machine Learning

Project Information

Link to Latest Report: Coming soon


Prediction of bridge condition based on bridge characteristics has been a topic of great interest to FHWA, state DoT’s, and researchers. Several methods for predicting bridge condition from the inventory information in the NBI are in the literature (e.g., Martinez et al. 2020, Cattan and Mohammadi, 1997, Rafiq et al., 2015, and Jiang, 2010, among other). While different approaches have been used to predict bridge condition, in general, all of them predict bridge condition more accurately with additional information on the individual bridges. Often collecting this additional information requires on-site bridge inspection, detailed review of bridge plans, and maintenance of large databases of this bridge information. Some Departments of Transportation (DOTs) are able to collect and maintain this more detailed information on
individual bridges. However, for many DOTs staffing levels are inadequate to support database maintenance at a high level.

With machine learning, it is possible to predict the detailed characteristics of bridge inventories at the individual bridge level using smaller training samples of bridges. The predicted inventory characteristics would use the NBI information on each bridge as input into a machine learning derived model that then predicts the more detailed characteristics of each bridge. These predicted inventory characteristics can then be used in existing models that predict, for example, long-term bridge condition, bridge seismic performance, and bridge performance in flooding and storm surge. Of course, such information is of great importance for DOT planning tasks, such as budget planning for maintenance and replacement, retrofit for functionality, and emergency operations planning in the event of natural hazards. While not as accurate as characteristics derived from detailed site inspections and plan review, the machine learning generated characteristics should provide a great improvement in the understanding of bridge inventory characteristics relative to NBI data alone.


The proposed project will break new ground in asset management for DOTs, will impact planning for bridge repair and replacement, and will provide critical inventory information for emergency planning in the face of extreme events such as natural hazards. The project will build upon an extensive database of bridge properties collected for more than 700 bridges in Washington State, leveraging substantial efforts from previous research projects that were supported by WSDOT, PacTrans, and the National Science Foundation. The detailed information on these 700 bridges is substantially more comprehensive than that in the NBI and can be used to select and apply models for prediction of long-term condition, seismic fragility, and storm surge and flooding fragility. Of particular importance, the data collected includes information on the substructure, including the piers, abutments, bearings, shear keys, and foundations, which are often critically important when predicting long-term bridge condition and bridge performance in natural hazards. Coupled with the NBI information and geographic information that provides the setting for the bridge (e..g., size of the river crossing), this is a unique data set that is available to the research team. The following tasks outline a year-long research project to utilize the collected data to train and assess the accuracy of a machine learning-based model to predict the characteristics of the remaining 4,000 plus bridges in Washington State, publish the model and prepare it for use in other states.


Task 1 – Train a Machine Learning Model:

Prof. Choe is a data scientist who specializes in machine learning with heterogenous data of infrastructures and natural hazards and will lead the development of a machine learning model for predicting key bridge characteristics using the data available in NBI and limited geographic setting information. The NBI information for the bridges in the detailed database has also already been curated. Prof. Choe and his student will develop the machine learning model in Python such that it is open source and can be shared and advanced in future projects. The modeling process will be multi-pronged, including highly interpretable models and latest deep learning architectures with few-shot tuning to obviate the need for a large training sample.

Task 2 –  Collect Detailed Information on Additional Bridges

: Extending the database of detailed WSDOT bridge characteristics will be completed by Prof’s. Berman, Eberhard and their student.The research team has access to drawings of the vast majority of bridges in WSDOT’s inventory and will use the drawings to collect the additional detailed information on at least 300 more bridges. The bridges will be systematically selected from the WSDOT inventory to maximize the database’s representation of the inventory in terms of NBI-based characteristics using a combination of representation learning and outlier ranking implemented by Prof. Choe.

Task 3 – Test the Bridge Inventory Characteristics Model

The machine learning derived predictive model will be used to predict the characteristics of the bridges in a representative hold-out dataset including some bridges databased in Task 2 using only their NBI information and geographic setting information. The model’s predictive performance will be analyzed to
inform the necessary revisions.

Task 4 – Model Revision if Necessary

If the results in Task 3 indicate opportunities to improve the model, Tasks 1-3 will be iterated until the revised model attains a satisfactory performance (i.e., reliably deployable in practice) on a representative test data set (never used in training).

Task 5 – Dissemination

The research team will disseminate the inventory prediction model through journal papers and presentation at national conferences. The team will prepare a report on the model development and accuracy as is standard for IBT/ABC UTC projects.

Research Team:
Principal Investigator:  Jeffrey Berman
Co-Principal Investigator: Marc Eberhard, Youngjun Choe