Abstract:
Fire-induced spalling is the phenomenon where the outer cracked or delaminated layer of a
concrete element detaches due to the exposure to high temperatures during a fire. Spalling is
a phenomenon that has raised concerns in the research community since the 19th century. Since
then, many experimental, analytical, numerical, and other studies have been conducted around
the world to explain this phenomenon. However, an accurate model to predict the occurrence
of spalling remains elusive, particularly for tunnel linings.
Tunnel fires have drawn increasing attention and raised more concerns in recent decades. The
rapid growth of freight transportation, particularly flammable ones such as fuel, increases the
potential to cause a rapid-fire spread. When compared to building fires, tunnel fires can be
more destructive due to their high temperatures, quick heating rates, prolonged duration, and
uneven temperature distribution inside the tunnel.
Spalling is a complex phenomenon with a high degree of randomness that interdepends on too
many factors. The occurrence of spalling phenomena is significantly influenced by various
microstructural properties of concrete. Internal factors such as concrete permeability, moisture
content, water-cement ratio, and aggregate type have a significant impact on spalling.
Furthermore, temperature, heating rate, humidity, and loading conditions are some of the
external factors that affect spalling. To gain a comprehensive understanding of this
phenomenon, it is crucial to consider the interdependencies among these various factors and
their combined effects.
The current method used in industry to evaluate the performance of a concrete tunnel lining is
to test the specimen in large-scale furnaces. However, this method has several limitations. It
requires the use of large-scale furnaces, which is time-consuming, expensive, and difficult to
replicate due to their dependence on specific concrete mixtures and test setups.
Alternative approaches, such as Machine Learning (ML), can be considered to overcome these
challenges. Recent advancements in data analytics & ML have demonstrated their capability
to solve such complex problems. This study aims to create a framework for predicting fireinduced
spalling in tunnel linings using several ML techniques.
Python programming language was utilised to develop this framework and Jupyter Notebook
was used as the web based interactive platform. Using the previously published fire test data,
a new dataset was created, and after performing the appropriate preprocessing, it was fed into
10 distinct ML techniques. These includes 7 ensemble techniques and 3 traditional ML
techniques. Then the developed model was further refined using hyperparameter tuning & kfold
cross-validation techniques. The results of this model revealed that it is possible to forecast
the occurrence of spalling with an accuracy of more than 90% using ensemble ML techniques.
