Mikiyasu ONDA

Efficiency Enhancement in Airborne Salt Particle Analysis through the Integration of Machine Learning and Numerical Simulation Technologies

Fuminori NAKAMURA

In coastal concrete structures, salt-induced deterioration from airborne chlorides is a severe issue, making salt environment prediction a vital countermeasure. While previous studies using 3D numerical simulations achieved high prediction accuracy for wind conditions and chloride transport, these advanced simulations require enormous computational time for long-term, multivariate environmental analyses. To resolve this, this study focuses on the Fourier Neural Operator (FNO), a machine learning method proven to streamline predictive analysis. We examined its applicability for predicting wind conditions near concrete structures and subsequently conducted a predictive analysis of airborne chlorides acting on the structures.
The machine learning model for wind prediction was trained using numerical simulation results. For airborne chloride prediction, the transport and arrival of chloride particles were simulated via the particle tracking method, utilizing the machine learning-predicted wind field as boundary conditions. The investigated cases included a fixed wind direction scenario reproducing wind tunnel experiments and a scenario with varying wind directions acting on the structure.
The results demonstrated that the machine learning method reduced the wind prediction time to approximately 10 seconds per case. The predicted wind speed distribution closely matched the numerical simulation results, with a relative error of only about 6%. Furthermore, the airborne chloride arrival process calculated using the predicted wind field aligned well with past wind tunnel experimental trends, proving the applicability of this method to salt environment prediction. Notably, under varying wind directions, the model successfully reproduced airflow changes corresponding to the structure's shape, even for untrained wind directions. This allowed for the accurate prediction of chloride particle arrival distributions based on wind direction.
In conclusion, the constructed machine learning model demonstrates the potential to efficiently predict and evaluate complex salt-induced environmental actions on concrete structures while drastically reducing the immense computational costs of conventional methods.