Civil infrastructure that supports our society faces risks from aging and natural disasters. Associate Professor Nishio and her team aim to make the management of infrastructure structures such as bridges more intelligent by leveraging advanced sensing technologies and numerical simulation.

Integrating Sensing Data into Numerical Simulations

Numerical simulation and numerical analysis are techniques for mathematically approximating physical phenomena and solving them using computers. In engineering, however, real-world objects and phenomena must be appropriately approximated and modeled, and the resulting problems must be solved quickly enough to meet decision-making needs. When numerical analysis is used to evaluate the safety of civil infrastructure structures such as bridges under loads encountered in real-world environments, including earthquakes and traffic, a variety of challenges need to be addressed.
One approach that has been actively studied in recent years is data assimilation. This technique reduces uncertainty in models used for numerical simulations by incorporating observational data, thereby improving the accuracy of simulation results. Through data assimilation, we aim to develop models that appropriately reflect the deterioration and damage of aging infrastructure. For example, we hope to develop technologies that can quantitatively assess whether the current condition of a structure is increasing its risk of damage from a possible future earthquake.
Data assimilation updates simulation models so that their outputs become consistent with observational data, using probabilistic estimation methods. Observational data obtained from large civil structures mainly concern deformation and vibration, and a wide range of sensing and information and communication technologies are used to collect these data. One of the interesting aspects of this research is the opportunity to use and think across different technological fields.
In recent years, installing or attaching sensors to large structures such as bridges has become a major practical challenge. We are therefore also studying data acquisition using image-based measurement and computer vision technologies. In addition, with the aim of enabling numerical simulations to be used rapidly for real-world problems, we have begun research on data-assimilation-based structural analysis combined with augmented reality (AR) technology.

Fig.1: Safety Analysis of Infrastructure Structures Using Data Assimilation
Data on target structures are acquired in the real world using various sensing technologies and incorporated into simulation models in the digital space for safety analysis. Machine learning and AI methods are used at various stages to support rapid decision-making in infrastructure management. (Figure by NISHIO Mayuko)

Improving the Efficiency of Numerical Simulations with Machine Learning

Alongside data assimilation, one of our current priorities is improving the efficiency of numerical simulations. We are exploring the use of machine learning and AI—in other words, replacing numerical analysis with machine-learning-based approaches. Although this line of research has gained momentum in many fields, not only engineering, our distinctive focus is on incorporating expertise in civil engineering and structural mechanics.
In general, developing a good AI model requires a large and diverse set of training data. Even when developing an AI model specialized for a particular purpose, such as structural analysis, producing a sufficient amount of training data can be very costly. To address this challenge, methods such as transfer learning can make machine-learning-based numerical analysis more efficient.
For example, when designing a bridge, we can first train a machine-learning model using analysis results for the bridge in its new, undamaged condition. Later, even if only a limited number of analyses can be performed for the bridge in a deteriorated or damaged condition, these results can be added as training data to enable numerical analysis of the damaged state. The role of transfer learning is to first learn the general characteristics of how a bridge responds as a structure, regardless of whether it is damaged.
A simple analogy would be an AI model designed to recognize different breeds of dogs from images. The model could first learn common features of animal faces—such as two eyes and one nose—from images of cats, and then apply that knowledge to learning about dogs.

Towards Handling More Complex Structural Systems

Data assimilation involves both numerical simulation and sensing. Laboratory studies often focus on relatively simple structures such as beams and plates, but what we ultimately have in mind are real-world structures. The bridges and buildings around us are composed of numerous structural members, forming highly complex systems. How to apply data assimilation to such complex systems is one of the challenges we intend to tackle next.

Even in numerical analysis alone, solving coupled physical equations across an enormous number of elements and locations requires substantial computational resources. Moreover, when we actually collect sensing data from structures in real-world environments, we obtain vast amounts of data containing complex noise that is far from ideal.
What makes these challenges particularly interesting is that addressing them requires us to look beyond structural engineering and learn from fields such as information science and applied mathematics. I enjoy studying new ideas across disciplinary boundaries and thinking about how they can be applied to our research.
I strongly feel that future engineering research must move beyond existing disciplinary frameworks, incorporate perspectives and methods suited to the objects and problems being studied, and create an environment in which such approaches are accepted.

Fig.2: Research on Data Assimilation Analysis of Structures
Data assimilation research requires not only numerical simulations but also the acquisition of measurement data, so validation work ranges widely from computer-based analysis to laboratory experiments and field measurements. The image on the left shows a beam vibration test conducted in the laboratory as part of research in which an AI-based structural vibration simulation is updated through real-time data assimilation and visualized using augmented reality (AR). The center image shows a bending test of a corroded beam, conducted to investigate whether actual deformation can be reproduced by representing the corrosion in a structural analysis model. The image on the right shows members of the laboratory acquiring three-dimensional data from an actual bridge.
Left image by Toko Okuda (AIS Nishio Lab, completed March 2025); center and right photos by members of the Nishio Lab.

(Date of interview: July 7, 2025)