炎熱天氣混凝土施工的機器學習：預測新拌及早齡期行為的批判性綜述

Oct 26, 2026 · @Ftohi@123
Abstract
Concreting in hot-arid climates accelerates slump loss, shortens setting time, raises evaporation from fresh surfaces and amplifies early-age thermal gradients. These effects are governed by interacting mixture, environmental and operational variables that conventional nomographs and empirical equations capture only partially. Machine learning (ML) has been applied extensively to hardened concrete, yet its use for the fresh and early-age behaviour that dominates hot-weather practice remains fragmented across materials, structural and construction-management literatures. This paper presents the first critical review dedicated to ML for hot-weather concreting. Peer-reviewed studies published in Scopus-indexed journals (quartiles Q1–Q4) were screened and organised around five phenomena: workability and slump loss, setting and hydration kinetics, evaporation and plastic shrinkage cracking, early-age temperature rise and thermal cracking, and in-place strength development. For each, the review compares algorithms, input features, data provenance and validation practice. The analysis shows that tree-based ensembles and neural networks achieve high accuracy on laboratory and simulated data, but few models are trained on site-measured hot-weather data, environmental variables are often aggregated or omitted, and external validation is rare. Physics-informed learning, transfer learning, explainability and sensor-driven digital twins are identified as the most promising routes to field-ready tools. A research roadmap is proposed for decision-support systems that help site engineers in hot regions schedule pours, dose retarders and select curing regimes.
Keywords: hot-weather concreting; machine learning; fresh concrete; plastic shrinkage; early-age thermal cracking; setting time; construction engineering

1. Introduction Concrete placed in hot weather behaves differently from concrete placed under laboratory conditions. High air and material temperatures, low relative humidity, wind and solar radiation increase water demand, accelerate slump loss and setting, and raise the rate of evaporation from exposed surfaces. Field studies in the Arabian Gulf showed that exposure conditions at the time of casting significantly affect plastic shrinkage, while elevated temperature also changes porosity, compressive strength and pulse velocity of the hardened material (Almusallam, 2001). Shrinkage of plain and silica-fume concretes is likewise amplified under hot-weather exposure (Al-Amoudi et al., 2007), and the response depends on cement type and water–cement ratio (Kockal & Türker, 2007). The construction consequences are practical and costly: rejected truckloads, cold joints, surface cracking, reduced durability and schedule disruption. Site decisions in hot weather still rely largely on prescriptive limits and empirical tools. The evaporation-rate nomograph and its closed-form equivalents estimate surface evaporation from air temperature, concrete temperature, humidity and wind speed (Uno, 1998). The maturity method relates temperature history to strength gain. These tools are valuable but simplified: they treat variables independently, ignore mixture-specific bleeding and setting behaviour, and cannot learn from the large volumes of batching, delivery and sensor data that modern projects now generate. Machine learning offers a way to model these non-linear, multivariate interactions. ML is now well established for predicting hardened properties, especially compressive strength (Ben Chaabene et al., 2020; Nunez et al., 2021; Young et al., 2019), and for mixture optimisation (DeRousseau et al., 2018). Concrete exposed to fire and elevated service temperatures has also received considerable ML attention. However, hot-weather concreting is a distinct problem: it concerns the first hours and days after mixing, under ambient rather than fire temperatures, and its outputs feed construction decisions rather than material design alone. Earlier reviews help position the gap. Gamil (2023) surveyed applications and trends of ML across concrete technology, and Adsul et al. (2025) reviewed numerical, code-based and ML models for concrete properties from the fresh to the long-term state. Neither treats the hot-weather construction setting, in which environmental time histories, haul and placement operations, and site decisions are central. The predictive toolkit itself is mature: artificial neural networks were applied to high-performance concrete strength more than two decades ago (Yeh, 1998), and gradient boosting and hybrid ensembles now achieve high accuracy on strength data (Kaloop et al., 2020; Asteris et al., 2021). The open question is how this toolkit has been, and should be, applied to the first hours and days of concrete in hot climates. No review has yet consolidated ML research on this construction-phase problem. Relevant studies are scattered across materials science (setting, hydration, shrinkage), structural engineering (thermal cracking of massive members) and construction informatics (sensing and monitoring). This fragmentation makes it hard for researchers to see what is solved, and for practitioners in hot regions, including the rapidly expanding construction sector of Saudi Arabia and the wider Gulf, to adopt ML tools with confidence. This review addresses that gap. Its objectives are to:
2. Map ML applications to the five phenomena that govern hot-weather concreting performance.
3. Compare algorithms, input variables, data sources and validation practices across studies.
4. Critically assess the readiness of current models for field decision-making in hot-arid climates.
5. Propose a research roadmap toward reliable, explainable and site-deployable tools. The remainder of the paper is organised as follows. Section 2 describes the review protocol. Section 3 summarises the physical mechanisms. Section 4 reviews ML applications by phenomenon. Section 5 examines data sources. Section 6 provides a critical analysis, Section 7 sets out research gaps and a roadmap, and Section 8 concludes.
6. Review methodology This paper is a structured critical review rather than a systematic review with meta-analysis. Its aim is to map and critically assess a fragmented, emerging literature, for which a narrative synthesis with transparent search and selection rules is more appropriate than exhaustive record counting. Reporting follows the principles of the PRISMA 2020 statement where they apply to this design (Page et al., 2021), in particular explicit sources, eligibility criteria and synthesis categories. Sources and search strategy. Literature was identified in 2026 through iterative keyword searches of academic search engines and publisher platforms, including ScienceDirect, SpringerLink, MDPI, the ICE Virtual Library and the ACI digital library. Searches combined three groups of terms:
7. Hot-weather and early-age terms: "hot weather", "hot climate", "plastic shrinkage", "slump loss", "setting time", "hydration heat", "thermal cracking", "maturity".
8. Material terms: concrete, cement paste, mortar.
9. Method terms: "machine learning", "neural network", "genetic algorithm", "random forest", XGBoost, "physics-informed". Backward and forward citation tracking (snowballing) from the most relevant studies located further work, including earlier hot-weather studies that keyword searches alone missed. Eligibility criteria. Studies were included if they were peer-reviewed journal articles in English, published in journals indexed in Scopus and ranked Q1–Q4 in the SCImago Journal Rank, and met at least one of three conditions: • They develop or evaluate an ML or soft-computing model for a fresh or early-age property, or for an environmental effect relevant to hot-weather placement and curing. • They provide experimental evidence on hot-weather effects that defines the inputs such models need. • They establish methods — physics-informed learning, explainability, validation practice, digital twins — needed to assess model readiness. Conference papers, theses, patents, trade literature and preprints were excluded, as were studies limited to fire or post-fire temperatures. Studies predicting 28-day strength from mixture proportions alone were retained only as methodological benchmarks. Screening and data extraction. Titles and abstracts were screened for relevance, followed by the full text where available. For each ML study the following were extracted: target property, algorithm, input variables, presence of environmental inputs, data source, dataset size, validation scheme and interpretability method. Studies were then grouped by the phenomenon they predict (Section 4). Composition of the reviewed corpus. The final corpus comprises 60 journal articles in four groups. Group Role in the review Articles ML models of fresh, early-age and hot-weather behaviour Core evidence synthesised in Section 4 19 Experimental and physical studies of hot-weather concreting Define model inputs and regional context (Sections 3 and 6) 18 ML in concrete engineering more broadly (strength models, prior reviews, explainability, transfer learning) Benchmarks and context 16 Methodological references (physics-informed learning, explainability, validation, digital twins, reporting) Criteria for assessing readiness 7 Total 60 Extending the venue criterion to Q3 and Q4 was deliberate: several early and directly relevant hot-weather studies appear in lower-quartile journals, and excluding them would misrepresent the field. Their spread across quartiles is itself evidence of the field's immaturity and is revisited in Section 6. Limitations of the approach. Because the search was iterative rather than a single fixed database query, the review does not claim exhaustive coverage or report record counts at each screening stage. Studies in languages other than English and industry grey literature were not included. These limits are typical of critical reviews of emerging topics and are partly offset by citation tracking and explicit eligibility rules.
10. Physical background of hot-weather concreting ML models are only as useful as the variables they are given. This section summarises the mechanisms that any predictive model for hot-weather concreting must represent, and identifies the input features they imply. 3.1 Workability and slump loss Higher material and ambient temperatures accelerate early hydration and water evaporation during mixing, transport and waiting time. The result is faster slump loss and greater water demand, which tempts crews to retemper the mix on site. Fresh concrete temperature also influences the strength eventually achieved, together with mixing time (Mahboub & Cutshaw, 2001). Relevant model inputs include concrete and air temperature at batching and discharge, elapsed time, drum revolutions, admixture type and dosage, and cement fineness. 3.2 Setting and hydration kinetics Hydration accelerates markedly as temperature rises (Schindler, 2004), so initial and final set occur earlier in hot weather. Early setting reduces the window for placement, consolidation and finishing, and increases the risk of cold joints. Setting time depends on cement chemistry and fineness, supplementary cementitious materials (SCMs), retarders and the full temperature history, not only the temperature at mixing. 3.3 Evaporation and plastic shrinkage cracking Plastic shrinkage cracking (PShC) develops in the first hours, while concrete is still plastic. Once bleed water has evaporated, capillary pressure in the near-surface pores causes contraction; if restrained, cracks form. Evaporation rate is commonly estimated from air temperature, concrete temperature, relative humidity and wind speed using the closed-form version of the nomograph (Uno, 1998). However, evaporation alone does not determine cracking: risk rises once evaporation outpaces bleeding and persists until the concrete gains tensile capacity, so bleeding and setting behaviour must be considered together with evaporation. Exposure conditions at casting have been shown to dominate plastic shrinkage under Gulf conditions (Almusallam, 2001). 3.4 Early-age temperature rise and thermal cracking In massive elements such as rafts, piers and abutments, heat of hydration produces a hot core and a cooler surface. The resulting gradients generate tensile stresses that can exceed early-age tensile strength. In hot climates, high placing temperatures raise the peak temperature and can increase the core–surface differential, especially when nights are cool. Thermal cracking risk depends on element geometry, binder content and type, placing temperature, ambient temperature cycles, formwork and insulation, as well as the thermal properties of the concrete itself (Do & Le, 2024; Klemczak et al., 2025). 3.5 Curing and in-place strength development Hot, dry, windy conditions remove moisture needed for hydration. Inadequate curing reduces strength, raises permeability and increases drying shrinkage (Al-Amoudi et al., 2007). The maturity concept links temperature history to strength gain, but standard maturity functions are calibrated for moderate temperatures and assume moisture is available. These assumptions are frequently violated on hot sites. 3.6 Field evidence from hot climates Experimental programmes carried out in or for hot regions quantify the variables an ML model must learn. Ortiz et al. (2005) simulated hot- and cold-weather thermal cycles and found that compressive strength was best when the difference between ambient and concrete temperature was smallest, which in hot weather meant mixing later in the day. Follow-up work on ready-mixed concrete identified aggregates as especially temperature-sensitive (Ortiz Lozano et al., 2008). Earlier studies examined mixing and placing practices in hot weather (Kayyali, 1984), the use of admixtures (Soroka & Ravina, 1998), the effect of retempering ready-mixed concrete in hot-dry environments (Alhozaimy, 2007) and the importance of the water–cement ratio for strength in hot climates (Ait-Aider et al., 2007). A sequence of Saudi studies examined casting temperature directly. Specimens cast at 25, 32, 38 and 45 °C and cured outdoors under wet burlap showed an optimum casting temperature of 32 °C for plain and silica-fume concretes and 38 °C for fly ash, very fine fly ash, slag and natural-pozzolan concretes (Nasir et al., 2016). Casting under natural summer conditions confirmed that blended concretes outperform plain concrete at later ages, a finding at odds with the common 35 °C placement limit (Khan & Nasir, 2021). Casting temperature and curing regime also jointly control strength and drying shrinkage of blended mixes (Nasir et al., 2022), and the choice between wet burlap and water- or acrylic-based curing compounds affects plastic and drying shrinkage (Al-Gahtani, 2010). In a different climate, Ambroziak and Ziolkowski (2020) tracked how daily temperature and rainfall during production, transport and placement shaped the strength of ready-mixed concrete delivered to a building site. These studies supply exactly the controlled variation — casting temperature, curing regime, binder type, time of day — that ML models need. Yet they are rarely compiled into training datasets. 3.7 Machine learning methods in brief Four families of methods recur in this literature. Artificial neural networks were the first to be applied to concrete (Yeh, 1998) and remain common for small datasets with smooth non-linear responses. Tree-based ensembles — random forests, gradient boosting and XGBoost — now dominate tabular prediction because they handle mixed variable types and need little tuning; gradient tree boosting (Kaloop et al., 2020) and hybrid ensembles of surrogate models (Asteris et al., 2021) illustrate their accuracy on strength data. Evolutionary methods, notably genetic algorithms and genetic programming, produce explicit equations that engineers can inspect and use without software. Physics-informed neural networks embed governing equations, such as heat conduction and hydration kinetics, directly in the training loss (Raissi et al., 2019). They belong to a broader physics-informed machine learning paradigm that is especially valuable when data are scarce (Karniadakis et al., 2021). Convolutional neural networks process images and are the standard tool for automated crack detection on concrete surfaces (Cha et al., 2017). Interpretability cuts across all families. Post-hoc tools such as SHAP decompose tree-model predictions into feature contributions at both local and global levels (Lundberg et al., 2020). An alternative position holds that, for high-stakes decisions, inherently interpretable models should be preferred to explaining black boxes after the fact (Rudin, 2019). Both views matter in construction, where a wrong prediction can mean cracked slabs or rejected loads. 3.8 Implications for model design Taken together, these mechanisms imply that a hot-weather ML model should combine three groups of inputs: • Mixture variables: binder content and type, SCMs, water–binder ratio, admixtures, aggregate properties. • Environmental variables: air and concrete temperature, relative humidity, wind speed, solar radiation, and their time histories. • Operational variables: batching-to-placement time, element geometry, curing method and timing, formwork type. Section 4 examines how far existing ML studies meet this requirement.
11. Machine learning applications by phenomenon The included studies are grouped by the phenomenon they predict. The central finding is uneven coverage: setting, hydration, shrinkage and thermal response are reasonably well served, whereas slump loss under heat and plastic shrinkage cracking — the two problems site engineers face most directly — have very few models, most of them outside the leading materials journals. 4.1 Workability and slump loss under heat Most ML workability studies predict slump from mixture proportions alone, treating it as a static property. Their environmental blind spot is significant, because slump in hot weather is a function of time and temperature. The most direct hot-weather application is that of Al-Khatib and Al-Martini (2019), who tested self-consolidating concrete outdoors during an Abu Dhabi summer at 25–40 °C, with mixing prolonged to 110 min. Their ANN models used ambient temperature, mixing time and SCM type as inputs to predict slump flow, relative yield stress and relative viscosity, and a Garson sensitivity analysis ranked input influence. Maqsoom et al. (2021) developed artificial neural network (ANN) and quadratic regression models for concrete cast and cured under hot-climate conditions, monitoring properties up to 180 days and recommending curing methods and water–cement ratios for site engineers. Vargas et al. (2024) built ML models that predict slump alongside compressive and flexural strength from mixture data, but without temperature inputs. General-purpose strength models such as adaptive boosting (Feng et al., 2020) show that ensemble learners handle non-linear mixture effects well, but they do not include temperature-dependent workability loss. At plant level, Chandwani et al. (2015) trained ANNs on ready-mix production records to predict slump and rank the relative importance of each constituent, showing that routine plant data can support learning. However, a model that predicts the full slump-loss curve from batching to discharge, conditioned on concrete and air temperature histories, has not yet been reported. A second hot-weather line of work used genetic algorithms to derive explicit rheological equations. Nehdi and Al-Martini (2009) combined oscillatory rheology with GA to estimate the yield stress of superplasticised cement paste as a function of ambient temperature and prolonged mixing, with computed values agreeing well with measurements. The approach was extended to concrete mixed for up to 110 min at high temperature, where yield stress varied linearly with elapsed time but followed power and inverse power functions of temperature and superplasticiser dosage (Al-Martini & Nehdi, 2009), and to GA-derived Bingham parameters of cement paste tested at 22–45 °C over 20–110 min (Al-Martini & Nehdi, 2010). Later work compared rheological models for superplasticised paste under high temperature and prolonged mixing (Al-Martini et al., 2021). Together these studies form the most systematic body of time–temperature workability data available, and their explicit equations could serve as physics priors for future ML models. 4.2 Setting time and hydration kinetics Setting is better served. Oey et al. (2020) trained ML estimators on a cement dataset to predict paste setting time and mortar strength development from cement composition and fineness. Guo et al. (2020) proposed a broad learning system that builds setting-time models rapidly from clinker composition and physical properties, aimed at industrial cement production. Lapeyre et al. (2021) used random forests to predict hydration kinetics of multicomponent cementitious systems, providing heat-release curves that underpin setting and temperature-rise predictions. A second line of work embeds physics into learning. Rahman et al. (2024) developed PINN-CHK, a physics-informed neural network that incorporates maturity functions and material properties to predict early-age temperature fields during cement hydration; they report a relative L2 error as low as 0.00341. These results matter for hot-weather practice because physics constraints reduce the data required, and data from hot sites are scarce. The limitation shared by these studies is that curing temperature is usually fixed or varied in controlled steps. None of the reviewed setting models was trained on fluctuating field temperature histories typical of a summer day in a hot-arid region. 4.3 Evaporation, plastic and early-age shrinkage Shrinkage prediction has advanced quickly. Hilloulin and Tran (2022) compiled a database on autogenous shrinkage of mixtures containing superabsorbent polymers (SAP) and SCMs; XGBoost was the most accurate of four models, and SHAP and partial dependence plots quantified input influence. Hilloulin and Umunnakwe (2024) extended the Bažant shrinkage dataset with modern SCM-rich, high-performance and ultra-high-performance concretes; an Extra Trees model reached R² = 0.957 for total shrinkage over time, and interpretability analysis ranked the most influential inputs. These models target autogenous and drying shrinkage at hardened ages. Plastic shrinkage cracking in the first hours is almost absent from the ML literature. Image-based ML has been used to estimate crack patterns and propagation on concrete surfaces (Bayar & Bilir, 2019), and convolutional neural networks detect cracks reliably in real images (Cha et al., 2017); together these suggest a path for automated PShC detection on site. On the environmental side, Papadimitropoulos et al. (2020) replaced the ACI 305R-10 evaporation nomograph with genetic-algorithm curve fitting and ANN models; GA-optimised fitting reproduced the nomograph most accurately, giving site engineers a fast evaporation estimate. However, no reviewed model predicts PShC risk itself by combining evaporation rate with bleeding and setting time. This is the single most conspicuous gap identified by the review. 4.4 Early-age temperature rise and thermal cracking Thermal response of massive elements is a mature target for ML because finite-element simulations can generate large training sets. Klemczak et al. (2025) compared ML models for predicting hydration-induced temperature rise and thermal gradients in mass concrete; XGBoost gave the highest accuracy for maximum temperature rise and core–surface temperature differential. Do and Le (2024) trained SVM, XGBoost, ANN and genetic programming models on 759 simulated bridge-pier cases, with ambient and initial concrete temperature among the inputs, to predict the maximum cracking index and the time of cracking; all models exceeded R² = 0.96, and genetic programming produced transparent symbolic equations usable in practice. Temperature measurements from real members are the benchmark such models need, but they are rarely used for training. Placing temperature and ambient temperature are standard inputs in this group, which makes it the most climate-aware branch of the literature. Yet most training data are simulated, and diurnal cycles characteristic of desert climates — very hot days followed by markedly cooler nights — are rarely represented explicitly. Monitoring data are beginning to enter this branch. Tyurina et al. (2026) trained ANNs on parametric finite-element results to predict thermal stress in mass-concrete foundation slabs from temperature readings at a few characteristic points, enabling real-time stress assessment at lower computational cost than conventional modelling. Such monitoring-driven surrogates are a natural bridge to the digital twins discussed in Section 5. 4.5 In-place strength development and thermal-control materials The maturity method is the industry standard for estimating in-place strength from temperature history, and commercial sensor platforms increasingly add ML-based projections. Despite this industrial uptake, peer-reviewed studies that train ML models on field maturity data under hot conditions remain scarce. Related work on materials for thermal control shows promise: Marani and Nehdi (2020) used ML to predict the compressive strength of cementitious composites incorporating phase change materials (PCMs), which are proposed to moderate temperature extremes in hot climates. Fresh-state measurements are also being used as model inputs. Zviazhynski et al. (2025) predicted concrete strength with machine learning using fresh-state measurements, an approach that could readily incorporate concrete temperature and slump at discharge on hot days. 4.6 Explainability across phenomena Explainable AI is increasingly reported alongside predictions. SHAP has been used to show that tree-based model explanations agree with known concrete behaviour (Ekanayake et al., 2022), and AutoML has been coupled with SHAP to automate model selection while retaining global and local explanations (Sun et al., 2023). For hot-weather applications, explainability is not optional: a site engineer must know whether a predicted crack risk is driven by wind, temperature or setting time in order to choose the right mitigation. 4.7 Comparative summary Study Phenomenon Algorithm(s) Main data source Explicit climate inputs Key result Al-Khatib & Al-Martini (2019) Slump flow and rheology of SCC under heat ANN, Garson sensitivity Outdoor tests, Abu Dhabi summer Yes (25–40 °C, mixing time) Predicts slump flow, yield stress and viscosity vs temperature and time Nehdi & Al-Martini (2009) Yield stress of paste under heat Genetic algorithms Oscillatory rheology tests Yes (ambient temperature, mixing time) Computed yield stress agreed well with measurements Chandwani et al. (2015) Slump of ready-mix concrete ANN Ready-mix plant records No Plant data usable; relative importance of constituents Maqsoom et al. (2021) Properties under hot-weather curing ANN, quadratic regression Laboratory and field-cured specimens Yes Models guide curing method and w/c choice in hot climate Oey et al. (2020) Setting, strength evolution Several ML estimators Cement dataset No (fixed conditions) Setting predicted from cement composition and fineness Guo et al. (2020) Setting time Broad learning system Industrial cement data No Fast, accurate setting-time models Lapeyre et al. (2021) Hydration kinetics Random forest Calorimetry, multicomponent binders Curing temperature Prompt prediction of heat-release curves Rahman et al. (2024) Hydration temperature field PINN Experimental + physics Temperature as state Relative L2 error 0.00341 Papadimitropoulos et al. (2020) Surface evaporation rate GA curve fitting, ANN ACI 305R-10 nomograph Yes GA fitting reproduced the nomograph most accurately Hilloulin & Tran (2022) Autogenous shrinkage XGBoost and others, SHAP Literature database Limited XGBoost most accurate Hilloulin & Umunnakwe (2024) Total shrinkage Extra Trees and other ensembles Extended Bažant database Test conditions R² = 0.957 Do & Le (2024) Thermal cracking of bridge piers SVM, XGBoost, ANN, GP 759 simulated cases Yes (ambient and initial concrete temperature) All models R² > 0.96; GP gives symbolic equations Klemczak et al. (2025) Mass-concrete temperature rise XGBoost and others Mass-concrete thermal cases Yes Best accuracy for peak rise and core–surface differential Tyurina et al. (2026) Thermal stress in mass-concrete slabs ANN Parametric FE results + monitoring points Temperature readings Real-time stress prediction from few sensors Marani & Nehdi (2020) Strength of PCM composites Several ML models Literature database No Supports PCM-based thermal control Two patterns stand out. First, ensemble trees (XGBoost, Extra Trees, random forest) dominate tabular problems, while physics-informed networks lead where spatial–temporal temperature fields matter. Second, explicit climate inputs are present mainly in thermal studies; fresh-state models rarely include humidity, wind or solar radiation.
12. Data sources and acquisition The accuracy reported by any ML model reflects the data it was trained on. Three data sources dominate the reviewed literature, each with distinct strengths and weaknesses for hot-weather applications. 5.1 Laboratory and literature databases Most models are trained on laboratory data, either generated by the authors or compiled from published studies. Compiled databases, such as the extended Bažant shrinkage dataset (Hilloulin & Umunnakwe, 2024) or multicomponent hydration data (Lapeyre et al., 2021), offer breadth and enable benchmarking. Their weakness for hot-weather work is that tests are run under controlled temperature and humidity. Environmental variables are therefore either constant or varied in a few discrete levels, which cannot represent a real summer pour. Compiled data also mix test methods and laboratories, introducing noise that models may learn as signal. 5.2 Numerical simulation Finite-element thermal and thermo-mechanical models can generate thousands of labelled cases cheaply, which explains why thermal cracking is the best-covered problem (Klemczak et al., 2025). Simulation also enables transfer learning, where a model is pre-trained on simulated data and fine-tuned on scarce real data. In structural engineering, transfer learning has improved prediction most when the target dataset is small (Pak & German Paal, 2022). The risk is that a surrogate inherits the simplifications of the simulator, such as idealised boundary conditions and constant ambient temperature. 5.3 Field sensing and production records Modern projects produce rich field data: batch tickets with material quantities and temperatures, truck telematics with drum revolutions and elapsed time, embedded temperature and humidity sensors, on-site weather stations, and image records of finished surfaces. Field records capture the real variability of hot-weather construction and are the natural basis for site decision tools. ML has a long record of turning monitoring data into durability and service-life assessments for reinforced concrete (Taffese & Sistonen, 2017). Yet, apart from outdoor testing such as that of Al-Khatib and Al-Martini (2019), field production and sensor data from hot-arid sites are almost never used to train fresh or early-age models. 5.4 Toward digital twins of the pour Combining these sources points toward a digital twin of the concrete pour, in the sense used for construction digital twins more broadly (Boje et al., 2020): a model that ingests batching data, live weather and embedded sensor readings, predicts setting, evaporation, temperature and strength in real time, and updates itself as measurements arrive. Physics-informed networks are well suited to this role because they respect hydration and heat-transfer laws while assimilating sparse sensor data (Rahman et al., 2024). No complete digital twin for hot-weather concreting has yet been reported in the reviewed literature, but its components exist.
13. Critical analysis High reported accuracy is common in the reviewed studies, but accuracy on a held-out laboratory split is not the same as reliability on a hot construction site. Six weaknesses limit the field readiness of current models. 6.1 Data scarcity in the hot-weather regime The conditions of greatest interest — concrete temperatures above about 32 °C, humidity below 30%, strong wind and intense solar radiation — sit at the edge of, or outside, most training sets. Tree-based models in particular cannot extrapolate: they return values bounded by the training range. A model trained under temperate conditions may therefore under-predict risk precisely when it matters most. Approaches that embed physical knowledge, such as physics-informed networks constrained by hydration and heat-transfer laws (Rahman et al., 2024), and transfer learning from simulation or related tasks (Pak & German Paal, 2022), are the most credible responses. 6.2 Missing and aggregated environmental variables Fresh-state models rarely include humidity, wind speed or solar radiation. Where environmental data are included, they are often averaged over a day, losing the hourly dynamics that control evaporation and setting. Since evaporation depends non-linearly on these variables (Uno, 1998), aggregation removes exactly the information a hot-weather model needs. 6.3 Weak validation practice Most studies use a single random train–test split from one dataset. Random splits leak information when several records come from the same mixture or the same specimen series, inflating apparent accuracy. External validation on an independent laboratory, plant or site is rare. Such leakage is a recognised source of irreproducible results across ML-based science (Kapoor & Narayanan, 2023). Very high R² values should therefore be read as measures of interpolation within one dataset, not as evidence of transferability. 6.4 Static outputs for a dynamic process Most models predict a single number — a slump, a setting time, a peak temperature. Site decisions, however, depend on trajectories: how fast slump is falling, when the evaporation rate will exceed a safe threshold, when the core–surface differential will peak. Sequence models and physics-informed approaches that predict full time histories are better aligned with construction decisions but remain few. 6.5 Interpretability and trust SHAP and related methods are increasingly reported (Ekanayake et al., 2022; Hilloulin & Tran, 2022; Sun et al., 2023), but explanations are usually global rankings of inputs. Site engineers need local, actionable explanations: for this pour, at this hour, which factor dominates the risk and what change would reduce it. Where stakes are high, inherently interpretable models may be preferable to post-hoc explanation (Rudin, 2019), while tree-explainer methods can provide consistent local attributions (Lundberg et al., 2020). Explanations must also be checked against physics; an explanation that contradicts known mechanisms signals a spurious model. 6.6 Publication venue and fragmentation The most directly relevant hot-weather studies — on slump flow and rheology under outdoor heat, plant-level slump prediction and evaporation-rate modelling (Al-Khatib & Al-Martini, 2019; Chandwani et al., 2015; Papadimitropoulos et al., 2020) — appear mostly outside the leading materials journals, which concentrate on hardened properties and simulation-based thermal studies. This reflects how recent and dispersed the topic still is. It also indicates an opportunity: rigorous hot-weather ML studies with field data would fill an evident gap in high-impact journals. 6.7 Regional considerations for Saudi Arabia and the Gulf Hot-arid regions add specific complications that general-purpose models ignore. • Binder-dependent temperature limits. Optimum casting temperatures differ by binder, with blended concretes tolerating higher values than plain concrete (Nasir et al., 2016; Khan & Nasir, 2021). A single temperature limit, or a single ML model trained mainly on plain mixes, may therefore be inappropriate. • Curing practice. Site curing varies widely, and curing compounds perform differently from wet curing (Al-Gahtani, 2010), yet curing method is rarely an input to ML models. • Retempering and label quality. Water added on site to restore workability of ready-mixed concrete in hot-dry conditions (Alhozaimy, 2007) changes the mixture actually placed. Unless delivery records capture it, such additions corrupt training labels. • Time of day. The strong diurnal cycle means that the hour of mixing affects performance (Ortiz et al., 2005), so models should include it explicitly rather than daily averages. Addressing these factors would make ML tools directly relevant to the large construction programmes now under way in the region.
14. Research gaps and roadmap The analysis points to seven research directions, ordered from the most urgent to the most ambitious. G1 — ML prediction of plastic shrinkage cracking risk. No reviewed model predicts PShC by combining evaporation rate, bleeding capacity and setting time. A hybrid model that couples data-driven evaporation estimates (Papadimitropoulos et al., 2020) with the physics of bleeding and setting, with ML learning mixture- and site-specific corrections, would directly support site decisions on evaporation retarders, windbreaks and fogging. G2 — Time-dependent slump-loss models conditioned on temperature. Models should predict the full slump trajectory from batching to discharge, using concrete and air temperature histories, haul time and admixture dosage. Outputs could drive retarder dosing at the plant. G3 — Open hot-weather datasets. The field needs shared, well-documented datasets from hot-arid sites, including batch tickets, hourly weather, embedded sensor records and outcome measurements. Gulf ready-mix producers and large public projects are well placed to provide them. G4 — Physics-informed and transfer learning for extrapolation. Because the critical conditions lie at the edge of available data, models must respect hydration kinetics, heat transfer and evaporation laws. PINNs (Rahman et al., 2024) and simulation-to-field transfer learning (Pak & German Paal, 2022) should be tested specifically under hot-weather extremes. G5 — Rigorous validation protocols. Studies should report grouped or leave-one-source-out cross-validation, external validation on independent sites, and uncertainty estimates, not only R² on random splits. G6 — Local, physics-consistent explanations. Explainability should move from global feature rankings to pour-specific explanations that a site engineer can act on, checked against known mechanisms (Ekanayake et al., 2022). G7 — Integrated decision-support and digital twins. The long-term goal is a pour-level digital twin that combines weather forecasts, batching data and embedded sensors to recommend placing time, retarder dosage, curing method and formwork removal time. Horizon Priority directions Expected deliverable Short term (1–2 years) G1, G2, G5 Validated PShC-risk and slump-loss models for hot climates Medium term (2–4 years) G3, G4, G6 Open regional datasets; physics-informed models that extrapolate reliably Long term (4+ years) G7 Site-deployed digital twin for hot-weather pours 7.1 A conceptual decision-support pipeline The directions above converge on a single workflow for a hot-weather pour.
15. Before batching: a weather forecast and the approved mix design feed models that predict slump loss, setting time and evaporation rate over the planned delivery and placement window, and recommend the mixing hour and retarder dosage.
16. During delivery: batch tickets, concrete temperature at discharge and elapsed time update the slump and setting predictions, flagging loads at risk of rejection or retempering.
17. During placement and finishing: on-site weather sensors and the evaporation model trigger mitigation, such as fogging, windbreaks or evaporation retarders, when PShC risk exceeds a threshold.
18. During curing: embedded temperature sensors feed a physics-informed model of the temperature field and in-place strength, guiding curing duration, insulation and formwork removal.
19. After the pour: measured outcomes return to the dataset, so the models improve with every project. Each step uses methods already demonstrated in isolation in the reviewed literature. What is missing is their integration, validation on hot-arid sites, and a shared data infrastructure.
20. Conclusions This review consolidated, for the first time, machine learning research relevant to hot-weather concreting and assessed its readiness for construction practice in hot-arid climates. The main conclusions are:
21. ML research on concrete remains concentrated on hardened properties; fresh and early-age behaviour under hot conditions is comparatively under-studied and scattered across disciplines.
22. Setting, hydration kinetics, shrinkage and thermal response of massive elements are reasonably well modelled, with ensemble trees leading tabular problems and physics-informed networks leading temperature-field problems.
23. Plastic shrinkage cracking and temperature-dependent slump loss — the problems site engineers face most directly in hot weather — have very few ML models, none validated on independent site data from hot-arid regions.
24. Most models are trained on laboratory or simulated data with constant or coarsely varied environmental conditions, are validated on random splits of a single dataset, and cannot be relied upon to extrapolate to extreme heat.
25. Physics-informed learning, transfer learning, local explainability and pour-level digital twins are the most promising routes to field-ready tools. For regions such as Saudi Arabia and the wider Gulf, where large construction programmes coincide with severe climate, closing these gaps would reduce cracking, rework and schedule risk. Regional experimental evidence on casting temperature, curing regime and binder type already exists for these conditions and should be compiled into shared training datasets. The roadmap proposed here offers researchers a structured path from isolated laboratory models to decision-support systems for hot-weather concreting. References Adsul, N., Choi, Y., & Kang, S.-T. (2025). A comprehensive review of numerical and machine learning approaches for predicting concrete properties: From fresh to long-term. Materials, 18(15), 3718. https://doi.org/10.3390/ma18153718 Ait-Aider, H., Hannachi, N. E., & Mouret, M. (2007). Importance of W/C ratio on compressive strength of concrete in hot climate conditions. Building and Environment, 42, 2461–2465. https://doi.org/10.1016/j.buildenv.2006.05.003 Al-Amoudi, O. S. B., Maslehuddin, M., Shameem, M., & Ibrahim, M. (2007). Shrinkage of plain and silica fume cement concrete under hot weather. Cement and Concrete Composites, 29(9), 690–699. Al-Gahtani, A. S. (2010). Effect of curing methods on the properties of plain and blended cement concretes. Construction and Building Materials, 24(3), 308–314. Al-Khatib, M. I., & Al-Martini, S. (2019). Predicting the rheology of self-consolidating concrete under hot weather. Proceedings of the Institution of Civil Engineers – Construction Materials, 172(5), 235–245. https://doi.org/10.1680/jcoma.16.00055 Al-Martini, S., Khartabil, A., & Nehdi, M. L. (2021). Estimating rheological properties of superplasticized cement paste at high temperature and prolonged mixing using various rheological models and oscillatory rheology. ACI Materials Journal, 118(6). Al-Martini, S., & Nehdi, M. (2009). Genetic algorithm-based yield stress equations for concrete at high temperature and prolonged mixing time. Computers and Concrete, 6(4), 343–356. Al-Martini, S., & Nehdi, M. (2010). Genetic algorithm rheological equations for cement paste. Proceedings of the Institution of Civil Engineers – Construction Materials, 163(2), 77–85. Alhozaimy, A. M. (2007). Effect of retempering on the compressive strength of ready-mixed concrete in hot-dry environments. Cement and Concrete Composites, 29, 124–127. https://doi.org/10.1016/j.cemconcomp.2006.08.007 Almusallam, A. A. (2001). Effect of environmental conditions on the properties of fresh and hardened concrete. Cement and Concrete Composites, 23(4–5), 353–361. Ambroziak, A., & Ziolkowski, P. (2020). Concrete compressive strength under changing environmental conditions during placement processes. Materials, 13(20), 4577. https://doi.org/10.3390/ma13204577 Asteris, P. G., Skentou, A. D., Bardhan, A., Samui, P., & Pilakoutas, K. (2021). Predicting concrete compressive strength using hybrid ensembling of surrogate machine learning models. Cement and Concrete Research, 145, 106449. Bayar, G., & Bilir, T. (2019). A novel study for the estimation of crack propagation in concrete using machine learning algorithms. Construction and Building Materials, 215, 670–685. https://doi.org/10.1016/j.conbuildmat.2019.04.227 Ben Chaabene, W., Flah, M., & Nehdi, M. L. (2020). Machine learning prediction of mechanical properties of concrete: Critical review. Construction and Building Materials, 260, 119889. https://doi.org/10.1016/j.conbuildmat.2020.119889 Boje, C., Guerriero, A., Kubicki, S., & Rezgui, Y. (2020). Towards a semantic construction digital twin: Directions for future research. Automation in Construction, 114, 103179. Cha, Y.-J., Choi, W., & Büyüköztürk, O. (2017). Deep learning-based crack damage detection using convolutional neural networks. Computer-Aided Civil and Infrastructure Engineering, 32(5), 361–378. Chandwani, V., Agrawal, V., & Nagar, R. (2015). Modeling slump of ready mix concrete using artificial neural network. International Journal of Technology, 6(2), 207–216. https://doi.org/10.14716/ijtech.v6i2.213 DeRousseau, M. A., Kasprzyk, J. R., & Srubar, W. V. (2018). Computational design optimization of concrete mixtures: A review. Cement and Concrete Research, 109, 42–53. Do, T. A., & Le, B.-A. (2024). Machine learning approach for predicting early-age thermal cracking potential in concrete bridge piers. Forces in Mechanics, 17, 100297. Ekanayake, I. U., Meddage, D. P. P., & Rathnayake, U. (2022). A novel approach to explain the black-box nature of machine learning in compressive strength predictions of concrete using Shapley additive explanations (SHAP). Case Studies in Construction Materials, 16. Feng, D. C., Liu, Z. T., Wang, X. D., Chen, Y., Chang, J. Q., Wei, D. F., & Jiang, Z. M. (2020). Machine learning-based compressive strength prediction for concrete: An adaptive boosting approach. Construction and Building Materials, 230, 117000. Gamil, Y. (2023). Machine learning in concrete technology: A review of current researches, trends, and applications. Frontiers in Built Environment, 9, 1145591. https://doi.org/10.3389/fbuil.2023.1145591 Guo, J., Wang, L., Fan, K., & Yang, B. (2020). An efficient model for predicting setting time of cement based on broad learning system. Applied Soft Computing, 96, 106698. https://doi.org/10.1016/j.asoc.2020.106698 Hilloulin, B., & Tran, V. Q. (2022). Using machine learning techniques for predicting autogenous shrinkage of concrete incorporating superabsorbent polymers and supplementary cementitious materials. Journal of Building Engineering, 49, 104086. Hilloulin, B., & Umunnakwe, R. (2024). Machine learning-aided prediction of shrinkage in modern concrete: Focus on mix proportions and SCMs. Journal of Building Engineering, 98, 111410. https://doi.org/10.1016/j.jobe.2024.111410 Kaloop, M. R., Kumar, D., Samui, P., Hu, J. W., & Kim, D. (2020). Compressive strength prediction of high-performance concrete using gradient tree boosting machine. Construction and Building Materials, 264, 120198. Kapoor, S., & Narayanan, A. (2023). Leakage and the reproducibility crisis in machine-learning-based science. Patterns, 4(9), 100804. Karniadakis, G. E., Kevrekidis, I. G., Lu, L., Perdikaris, P., Wang, S., & Yang, L. (2021). Physics-informed machine learning. Nature Reviews Physics, 3(6), 422–440. Kayyali, O. A. (1984). Effect of certain mixing and placing practices in hot weather on the strength of concrete. Building and Environment, 19(1), 59–63. https://doi.org/10.1016/0360-1323(84)90015-5 Khan, M. U., & Nasir, M. (2021). Influence of in-situ casting temperature and curing regime on the properties of blended cement concretes under hot climatic conditions. Construction and Building Materials, 272, 121865. Klemczak, B., Bąba, D., & Siddique, R. (2025). Machine learning-based prediction of heat transfer and hydration-induced temperature rise in mass concrete. Energies, 18(17), 4673. Kockal, N. U., & Türker, F. (2007). Effect of environmental conditions on the properties of concretes with different cement types. Construction and Building Materials, 21(3), 634–645. Lapeyre, J., Han, T., Wiles, B., Ma, H., Huang, J., Sant, G., & Kumar, A. (2021). Machine learning enables prompt prediction of hydration kinetics of multicomponent cementitious systems. Scientific Reports, 11, 3922. Lundberg, S. M., Erion, G., Chen, H., DeGrave, A., Prutkin, J. M., Nair, B., Katz, R., Himmelfarb, J., Bansal, N., & Lee, S.-I. (2020). From local explanations to global understanding with explainable AI for trees. Nature Machine Intelligence, 2(1), 56–67. Mahboub, K. C., & Cutshaw, Q. A. (2001). Effects of fresh concrete temperature and mixing time on compressive strength of concrete. ACI Materials Journal, 98. https://doi.org/10.14359/10161 Maqsoom, A., Aslam, B., Gul, M. E., Ullah, F., Kouzani, A. Z., Mahmud, M. A. P., & Nawaz, A. (2021). Using multivariate regression and ANN models to predict properties of concrete cured under hot weather: A case of Rawalpindi, Pakistan. Sustainability, 13(18), 10164. https://doi.org/10.3390/su131810164 Marani, A., & Nehdi, M. L. (2020). Machine learning prediction of compressive strength for phase change materials integrated cementitious composites. Construction and Building Materials, 265, 120286. Nasir, M., Adesina, A., Ibrahim, M., Khan, M. U., Al-Amoudi, O. S. B., & Ali, S. I. (2022). Role of casting and curing conditions on the strength and drying shrinkage of greener concrete. Environmental Science and Pollution Research, 29(48), 72598–72610. Nasir, M., Al-Amoudi, O. S. B., Al-Gahtani, H. J., & Maslehuddin, M. (2016). Effect of casting temperature on strength and density of plain and blended cement concretes prepared and cured under hot weather conditions. Construction and Building Materials, 112, 529–537. Nehdi, M., & Al-Martini, S. (2009). Estimating time and temperature dependent yield stress of cement paste using oscillatory rheology and genetic algorithms. Cement and Concrete Research, 39(11), 1007–1016. Nunez, I., Marani, A., Flah, M., & Nehdi, M. L. (2021). Estimating compressive strength of modern concrete mixtures using computational intelligence: A systematic review. Construction and Building Materials, 310, 125279. Oey, T., Jones, S., Bullard, J. W., & Sant, G. (2020). Machine learning can predict setting behavior and strength evolution of hydrating cement systems. Journal of the American Ceramic Society, 103(1), 480–490. https://doi.org/10.1111/jace.16706 Ortiz, J., Aguado, A., Agulló, L., & García, T. (2005). Influence of environmental temperatures on the concrete compressive strength: Simulation of hot and cold weather conditions. Cement and Concrete Research, 35(10), 1970–1979. Ortiz Lozano, J. A., Aguado de Cea, A., Agulló Fité, L., García Vicente, T., & Zermeño de León, M. E. (2008). Experimental study of the effect of temperature on the strength of ready-mixed concrete. Theory. Materiales de Construcción, 58(291), 7–22. https://doi.org/10.3989/mc.2008.v58.i291.103 Page, M. J., McKenzie, J. E., Bossuyt, P. M., Boutron, I., Hoffmann, T. C., Mulrow, C. D., et al. (2021). The PRISMA 2020 statement: An updated guideline for reporting systematic reviews. BMJ, 372, n71. Pak, H., & German Paal, S. (2022). Evaluation of transfer learning models for predicting the lateral strength of reinforced concrete columns. Engineering Structures, 266, 114579. Papadimitropoulos, V. C., Tsikas, P. K., & Chassiakos, A. P. (2020). Modeling the influence of environmental factors on concrete evaporation rate. Journal of Soft Computing in Civil Engineering, 4(4), 79–97. https://doi.org/10.22115/SCCE.2020.246071.1254 Rahman, M. A., Zhang, T., & Lu, Y. (2024). PINN-CHK: Physics-informed neural network for high-fidelity prediction of early-age cement hydration kinetics. Neural Computing and Applications, 36(22), 13665–13687. https://doi.org/10.1007/s00521-024-09791-y Raissi, M., Perdikaris, P., & Karniadakis, G. E. (2019). Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378, 686–707. Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1(5), 206–215. Schindler, A. K. (2004). Effect of temperature on hydration of cementitious materials. ACI Materials Journal, 101(1), 72–81. Soroka, I., & Ravina, D. (1998). Hot weather concreting with admixtures. Cement and Concrete Composites, 20, 129–136. https://doi.org/10.1016/S0958-9465(98)80005-X Sun, B., Cui, W., Liu, G., Zhou, B., & Zhao, W. (2023). A hybrid strategy of AutoML and SHAP for automated and explainable concrete strength prediction. Case Studies in Construction Materials, 19, e02405. Taffese, W. Z., & Sistonen, E. (2017). Machine learning for durability and service-life assessment of reinforced concrete structures: Recent advances and future directions. Automation in Construction, 77, 1–14. Tyurina, V. S., Chepurnenko, A. S., & Yazyev, B. M. (2026). Prediction of thermal stress in hardening mass concrete structures using temperature monitoring data. Structural Mechanics of Engineering Constructions and Buildings, 22(2), 93–104. Uno, P. J. (1998). Plastic shrinkage cracking and evaporation formulas. ACI Materials Journal, 95(4), 365–375. Vargas, J. F., Oviedo, A. I., Ortega, N. A., Orozco, E., Gómez, A., & Londoño, J. M. (2024). Machine-learning-based predictive models for compressive strength, flexural strength, and slump of concrete. Applied Sciences, 14(11), 4426. https://doi.org/10.3390/app14114426 Yeh, I.-C. (1998). Modeling of strength of high-performance concrete using artificial neural networks. Cement and Concrete Research, 28(12), 1797–1808. Young, B. A., Hall, A., Pilon, L., Gupta, P., & Sant, G. (2019). Can the compressive strength of concrete be estimated from knowledge of the mixture proportions? New insights from statistical analysis and machine learning methods. Cement and Concrete Research, 115, 379–388. Zviazhynski, B., White, C., Lees, J. M., & Conduit, G. J. (2025). Predicting concrete strength using machine learning and fresh-state measurements. Data-Centric Engineering, 6, e37. https://doi.org/10.1017/dce.2025.10018