AI-based cognitive framework for evaluating response of concrete structures in extreme conditions☆
Introduction
Concrete, an inert material, has superior properties which makes it well suited for use in extreme environments such as that associated with terrestrial (i.e. nuclear power plants) and extraterrestrial (i.e. lunar bases) applications where high temperatures and rapid temperature changes take place (Naser and Chehab, 2018, Takeuchi et al., 1998). Concrete maintains this superior behavior despite the fact that it undergoes a series of chemical and physio-mechanical changes that adversely affect its composition and nature. In some cases, these changes may alter key characteristics of concrete by developing cracks, inducing creep deformation or spalling (i.e. explosive reduction of cross-section driven by fire effects) (Kodur, 2014). As a result, predicting thermal and/or structural response of concrete structures concrete structural members/systems becomes a challenging task. This has been thoroughly documented over the past few decades (Huang et al., 1996, Naser, 2016a, Shakya and Kodur, 2017).
Early research on tracing fire response of concrete structures started by examining the performance of structural members and assemblies in specially designed furnaces. In these fire tests, a concrete element is exposed to a pre-defined “standard” temperature–time curve such as that of the ASTM E119 (ASTM E119, 2016) or ISO 834 (ISO 834- 1, 1999). In many instances, a tested element is loaded with a gravity loading corresponding to a portion of its sectional capacity i.e. 50% of that at ambient conditions. The fire-tested element could also be instrumented with thermocouples and deformation measuring devices to monitor its thermal and structural response during the fire. Once a fire test starts, the performance of the fire-tested element is closely monitored and documented. The fire test is terminated once the fire-weakened element exceeds a failure limit state, often when temperature at the unexposed side of the element or once deflection of element exceeds a predefined limit state. This point in time, when a structural element fails, is referred to as fire resistance.
Results from such fire tests were then compiled into tables, and then used to derive correlation equations that can estimate fire resistance of concrete elements (with features and conditions similar to those tested earlier). More recently, and due to the growing complexity of fire tests and lack of testing facilities, researchers and designers sought other means to evaluate fire response of concrete structures. With the advent of technological and computing advancement, the use of numerical techniques such as those associated with finite element (FE) analysis, has surged (Hawileh and Naser, 2012). While such techniques provide a suitable and, in a way, clean/affordable prediction of fire response of concrete structures, the lack of proper validation and standardization of solution process (i.e. solving algorithms), required inputs (e.g. material properties, heat transfer boundaries etc.), as well as need for special software (which often require special licenses, certified expertise and demanding computing resources) continue to hinder the application and acceptance of numerical techniques (Naser, 2016b).
From the vantage point of this work, most of the above challenges could be concurred through assimilating a new form of calculation techniques that leverages Artificial Intelligence (AI) to exploit relationships between key response parameters often linked with the fire problem or phenomenon. This is stems from the notion that AI has been widely used in a range of civil engineering sub-disciplines such as structural health monitoring (Naser and Kodur, 2018), transportation (Ledoux, 1997), seismic and wind design (Chiaruttini et al., 1989), material sciences (Naser, 2018), yet has not been fully incorporated into structural fire engineering and fire safety applications.
The use of artificial neural networks (ANNs) was also specifically applied towards concrete structures primarily to predict sectional capacity and/or structural response. In one study, Sanad and Saka (2001) investigated the use of ANNs to evaluate ultimate shear strength of reinforced-concrete deep beams by examining 111 data points. The outcome of this analysis shows that predictions from ANN can outperform that obtained from ACI codal provisions as well as Mau-Hsu method. Jadid and Fairbairn (1996) also outlined an ANN-based framework to predict moment curvature of concrete beams under ambient conditions. This framework was shown to be easy to adopt and most importantly of high accuracy. Ahmadi-Nedushan (2012) optimized instance-based learning approaches, compiled with generalized regression neural network and stepwise regressions, to predict the compressive strength of high performance concrete (HPC) given due consideration to mix proportion e.g. water to binder ratio, water/fly ash content among others. Hadi (2003) presented a comprehensive review on the application of ANNs into concrete structures, still, a thorough assessment of open literature shows that despite a handful of studies (Naser et al., 2012, Chan et al., 1998, McKinney and Ali, 2014, Lazarevska et al., 2014, Naser, 2019c, Naser, 2019b), the application of AI into structural fire engineering and fire safety continues to be lagging and did not reach its full potential yet.
In its simplest form, Chan et al. (1998) developed an ANN able of quantifying magnitude of degradation in concrete strength under elevated temperatures (up to 1200 C). This ANN was developed using published experimental data points and was then applied to estimate the degradation in compressive strength property of concrete made of varying mix proportions and exposed to different environmental factors. It is worth noting that the maximum prediction error between the developed ANN and the experimental results was less than 15%.
Few researchers were able of developing AI models capable of predicting other aspects of fire response/behavior of concrete structures. For example, McKinney and Ali (2014) developed a crude set of ANNs in order to qualitatively predict fire-induced spalling in concrete. These ANNs were first trained using actual observations from fire tests and then tested to validate their prediction capabilities. After 1500 training iterations, predictions obtained from these networks showed close agreement with that obtained from fire resistance tests and achieved a 0.5% error rate. Similarly, Lazarevska et al. (2014) also trained a fuzzy-based neural network (FNN) as to evaluate expected time to failure (i.e. fire rating) of RC elements. These researchers also noted the suitability of FNNs especially in cases where there is virtually insufficient experimental and/or numerically data available on fire response of concrete columns.
Naser et al. (2012) was also able to predict thermal response in insulated FRP-retrofitted T-shaped reinforced concrete (RC) beams through a newly developed ANN. In this work, these researchers examined the effect of fire intensity, type of insulation material and thickness on temperature rise in insulated and strengthened FRP-RC beams. The results from the developed ANN were arranged into design charts/aids that can be used to select appropriate insulation material and thickness for FRP-RC beams expected to be subjected to standard or realistic (design) fires. These design aids could help in practical situations and provide an easy-to-pick insulation scheme for fire conditions similar to that to occur in buildings. The same research group was also able of utilizing ANNs and genetic algorithms to derive material models for various construction materials including high strength, high performance and fiber-reinforced concretes (Naser, 2019c, Naser, 2019b). In a separate work, Erdem (2010) also trained an ANN to predict fire-caused loss experienced in flexural capacity of concrete slabs. The trained ANN can properly account for seven inputs, including concrete strength, reinforcement yield strength, effective depth of slab, and fire duration. Erdem (2010) used 294 data points and reported how the developed ANN is able of achieving high prediction capabilities with a correlation coefficient of 99.775% and 99.750% for training and testing, respectively.
Unlike previously published works, this study seeks the development of a cognitive approach that is based on symbolic regression and genetic algorithms as to realize the complex thermal and structural behavior of RC structural members exposed to extreme temperatures (exceeding 1200 °C). This framework has led to deriving simple expressions that are capable of evaluating temperature and deformation histories in a concrete member; at a specific point in time, or through tracing time–temperature/deformation history and up to four hours of exposure to a standard fire. These expressions are built to account for critical response parameters i.e. geometry of RC beams and columns, aggregate type used in concrete mix, steel reinforcement ratio, applied load level, thickness of concrete cover, fire exposure duration as well as compressive strength of concrete and yield strength of reinforcing steel. Furthermore, these expressions implicitly account for high temperature properties of concrete and reinforcing steel, as well as associated fire phenomena expected to occur in fire; such as creep and fire-induced spalling, and thus is not of need to collect or input of material properties nor acquiring special software for fire analysis.
In total, seven expressions were derived using the developed cognitive framework; two for evaluating thermal response (one for RC beams and one for RC columns) and five for evaluating structural response (one for RC beams and four for RC columns with varying aggregate types). The validity of the proposed simple expressions was cross-checked against fire-tested RC beams and columns collected from published works and open literature. The practical implications of integrating AI-based modeling, as well as applicability of extending derived expressions to RC beams and columns of various geometric properties, restraint conditions, and concrete strength classes is also discussed.
Section snippets
Behavior of reinforced concrete members under fire conditions
Before introducing the developed AI-based framework, a concise review of fire behavior of RC structural members is beneficial. This section highlights main mechanisms associated with thermal and structural behavior of RC beams and columns under fire conditions.
When a RC member is exposed to fire conditions, cross-sectional temperature in this member slowly rises. This slow rise in temperature arises from the good thermal (insulating) properties of concrete. Due to the presence of moisture, low
Artifical intelligence — background, rationale, and model development
In contrast to statistical approaches, AI does not involve assumptions to start examining a phenomenon. Instead, AI is a specially designed computational technique that hopes to replicate human-like thinking/cognition ability to solve complex engineering problems that may not be appropriately solved in a timely manner using conventional methods or would require complex solvers or environments (software). AI is suitable for engineering scenarios in which there is a large amount of inputs (random
Development of fire test databases
The first step towards developing fire test databases to serve as training and testing data points is to carry out a comprehensive review of open literature to pinpoint suitable studies/reports in which RC beams and columns were tested under standard fire conditions. This section covers selected fire tests and further presents insights into the development of AI framework. Full details on collected tests, together with material properties and loading conditions in each test, is spared herein
Performance and validation of AI-based derived expressions
Upon completion of collecting the above databases, these databases were input into the AI environment; developed by Searson (2009). In this software, candidate expressions are derived through symbolic regression to arrive at a relation between thermal and structural-based input parameters, i.e. fire exposure time (t), compressive strength of concrete () etc. Each relation encompasses a number of operators i.e. , −and/or, mathematical functions (sin, cos.). The compiled input parameters were
Extension of AI-derived expressions to new scenarios
Fig. 3, Fig. 4 show the validity of the proposed expressions when examined against 30% of the data points that were part of the developed databases. To further validate the predictability of the derived expressions in new scenarios to which they have not been seen before, these expressions were tested to examine the fire response of new RC beams and columns that were not part of the developed databases. At the time of this study, these scenarios include different cross-sectional sizes,
Practical implications, challenges, and future research
This study presents a concept that capitalizes on the fact that artificial intelligence (AI) and machine learning techniques seem to be rapidly evolving and this favors utilizing such methods into engineering applications. In this optimistic view, one must realize that the simplicity of AI-derived or AI-developed approaches does not come easily. As such, some of the challenges and limitations of utilizing AI into practical solutions and applications need to be highlighted. For a start, the
Conclusions
This work fosters artificial intelligence as a modern assessment tool into structural fire engineering and fire safety applications. This study highlights the development of a cognitive framework capable of predicting thermal and structural response of fire-exposed RC beams and columns as well as sheds light into some of the limitation and key challenges associated with incorporating AI into fire engineering and safety. The following key items could also be drawn from the results of this work:
- •
Conflict of interest
The authors declare that there is no conflict of interest in this paper.
References (51)
An optimized instance based learning algorithm for estimation of compressive strength of concrete
Eng. Appl. Artif. Intell.
(2012)- et al.
The experimental behavior of FRP-strengthened RC beams subjected to design fire exposure
Eng. Struct.
(2011) - et al.
Behavior of thermally restrained RC beams in case of fire
Eng. Struct.
(2018) - et al.
Experimental analysis on flexural behaviour of RC beams strengthened with CFRP laminates and under fire conditions
Compos. Struct.
(2018) - et al.
The structural behavior and simplified thermal analysis of normal-strength and high-strength concrete beams under fire
Eng. Struct.
(2011) Prediction of the moment capacity of reinforced concrete slabs in fire using artificial neural networks
Adv. Eng. Softw.
(2010)- et al.
Three-dimensional finite element modelling of the fire behaviour of insulated RC beams strengthened with EBR and NSM CFRP strips
Compos. Struct.
(2018) Neural networks applications in concrete structures
Comput. Struct.
(2003)- et al.
Thermal-stress analysis of RC beams reinforced with GFRP bars
Compos. Part B
(2012) Limits of spalling of fire-exposed concrete
Fire Saf. J.
(2003)
Non-linear finite element model to predict temperature histories within reinforced concrete in fires
Build. Environ.
Neural-network applications in predicting moment–curvature parameters from experimental data
Eng. Appl. Artif. Intell.
The performance of near-surface mounted CFRP strengthened RC beam in fire
Fire Saf. J.
An urban traffic flow model integrating neural networks
Transp. Res. Part C
Autonomous and resilient infrastructure with cognitive and self-deployable load-bearing structural components
Autom. Constr.
Properties and material models for modern construction materials at elevated temperatures
Comput. Mater. Sci.
Analysis of RC T-beams strengthened with CFRP plates under fire loading using ANN
Constr. Build. Mater.
Cognitive infrastructure-a modern concept for resilient performance under extreme events
Autom. Constr.
Experimental investigation on fire endurance of insulated concrete beams strengthened with near surface mounted FRP bar reinforcement
Compos. Part B
Fire resistance and spalling performance of confined concrete columns
Constr. Build. Mater.
Study on a concrete filled structure for nuclear power plants
Nucl. Eng. Des.
Fire resistance of reinforced concrete columns with L-, T-, and+-shaped cross-sections
Fire Saf. J.
Fire behavior of concrete T-beams strengthened with near-surface mounted FRP reinforcement
Eng, Struct.
Experimental study on the fire resistance of rc beams strengthened with near-surface-mounted high-Tg BFRP bars
Compos. Part B
Standard Methods of Fire Test of Building Construction and Materials
Cited by (40)
RAGN-L: A stacked ensemble learning technique for classification of Fire-Resistant columns
2024, Expert Systems with ApplicationsEvaluating the effect of curing conditions on the glass transition of the structural adhesive using conditional tabular generative adversarial networks
2024, Engineering Applications of Artificial IntelligenceRefined self-attention mechanism based real-time structural response prediction method under seismic action
2024, Engineering Applications of Artificial IntelligenceFAST-AlertNet: Early warning fire-induced collapse of large-span steel truss structures
2023, Engineering Applications of Artificial IntelligenceState-of-the-art AI-based computational analysis in civil engineering
2023, Journal of Industrial Information IntegrationOptimization of high-performance concrete mix ratio design using machine learning
2023, Engineering Applications of Artificial Intelligence
- ☆
No author associated with this paper has disclosed any potential or pertinent conflicts which may be perceived to have impending conflict with this work. For full disclosure statements refer to https://doi.org/10.1016/j.engappai.2019.03.004..