Experimental and ANN-Based Evaluation of Water-Based Al₂O₃, TiO₂, and CuO Nanofluids for Enhanced Engine Cooling Performance

Abstract

This study presents an integrated experimental and computational investigation into the thermal and hydraulic performance of three oxide-based nanofluids: aluminum oxide (Al₂O₃), titanium dioxide (TiO₂), and copper oxide (CuO) for advanced engine cooling applications. A custom-built test rig was used to assess nanofluid behavior under varying flow rates, nanoparticle volume fractions, and temperature gradients, replicating realistic engine conditions. According to the results, at ideal concentrations, CuO nanofluids continuously demonstrate better heat transfer properties, outperforming TiO₂ by up to 15% and AlO₃ by 7%. However, performance plateaus beyond 1.5% volume fraction due to increased viscosity and pressure drop. A multilayer feedforward artificial neural network (ANN) model was developed to predict convective heat transfer coefficients and friction factors based on experimental inputs, achieving a mean absolute percentage error below 5% and a coefficient of determination (R²) exceeding 0.98. The ANN demonstrated robust generalization across varying operating conditions and nanoparticle types, confirming its utility for surrogate modeling and optimization. This work is distinguished by its dual focus on thermal efficiency and hydraulic stability, as well as its use of data-driven modeling validated by empirical results. The findings provide actionable insights for thermal management system design in internal combustion, hybrid, and electric vehicles, where efficient, compact, and reliable cooling solutions are increasingly vital. The study advances the practical application of nanofluids by offering a comparative, ANN-validated framework that bridges the gap between lab-scale performance and real-world automotive cooling demands.

1. Introduction

As engine power densities increase and environmental regulations become more stringent, effective heat management has emerged as a critical challenge in contemporary internal combustion engine (ICE) systems. Cooling systems are under more stress as engine downsizing, fuel efficiency, and pollution reductions become more and more important. Conventional coolants such as water, ethylene glycol (EG), and water–EG mixtures are chemically stable but have poor thermal conductivity (typically below 0.6 W/m·K), limiting their ability to effectively remove heat under high thermal loads [1,2]. As a result, these fluids often require larger radiators and higher pumping power to maintain proper thermal regulation, which increases the weight and energy consumption of the cooling system. Several studies [3,4] have shown that improving coolant thermal performance can directly improve engine lifespan, fuel economy, and emission outcomes [5,6]. Therefore, developing innovative coolant formulations that enhance heat transfer while minimizing pressure drop is an urgent need in the field of automotive thermal management.

To overcome the natural limitations of base fluids, researchers have suggested using nanofluids, which are engineered colloidal suspensions of nanoparticles in traditional coolants. Adding high thermal conductivity nanoparticles like Al₂O₃, TiO₂, and CuO has been shown to greatly improve the fluid’s effective thermal conductivity, often by 10–40% even at low particle concentrations [7,8,9]. This thermal enhancement, caused by mechanisms such as Brownian motion, nanoparticle clustering, and micro-convection, boosts the heat transfer coefficient and allows for more efficient heat dissipation from engine parts. Al₂O₃ and TiO₂ nanoparticles are especially valued for their chemical inertness and cost-effectiveness, while CuO nanoparticles offer better thermal conductivity but tend to increase the fluid’s density and viscosity slightly [10,11]. Although many individual studies have shown these benefits in controlled laboratory settings, real-world automotive cooling systems face additional challenges like transient temperature changes, vibrations, and fluctuating flow rates, factors often missing in simplified tests. Still, the absence of comprehensive research combining experimental data with data-driven methods under engine-like conditions continues to hinder the practical use of nanofluids in vehicle systems [11,12].

Furthermore, it is crucial to understand that improving thermal conductivity alone is not enough to assess a nanofluid’s effectiveness in an engine cooling system. Pressure drop and viscosity increase are also important because they directly affect hydraulic stability and pumping power. Previous studies by [13,14,15] showed that some nanofluids, while improving thermal performance, could cause undesirable increases in flow resistance, which can reduce overall system efficiency. Therefore, an ideal coolant should balance thermal enhancement with hydraulic performance. In the current research, which typically concentrates on either heat transfer or fluid dynamics alone, these two aspects are rarely investigated together [16,17,18]. Additionally, empirical results often differ due to variations in nanoparticle size, shape, surfactants, and preparation methods, limiting the reproducibility and wider applicability of findings. The absence of a comprehensive thermal-hydraulic performance framework makes it challenging for engineers to make informed decisions about using nanofluids in real-world applications [19,20,21].

To address these gaps, a growing number of studies have turned to predictive modeling techniques to supplement experimental research and guide coolant selection. Among these, artificial neural networks (ANNs) have shown remarkable success in capturing the complex, nonlinear interactions between thermophysical properties and system-level performance parameters [22,23,24]. ANNs, when properly trained, can generalize across multiple operating conditions and nanoparticle types, thereby providing a powerful tool for predicting performance metrics such as the convective heat transfer coefficient, Nusselt number, and pressure drop. However, relatively few studies have applied ANN-based models specifically to the context of automotive cooling systems, and even fewer have validated these predictions against extensive experimental data under engine-like thermal and flow conditions. This limited integration of data-driven modeling with real-world testing represents a missed opportunity in the current literature.

This study proposes a comprehensive experimental and computational investigation to evaluate the thermal and hydraulic performance of three oxide-based nanofluids, Al₂O₃, TiO₂, and CuO, used as working fluids in an engine cooling system. Unlike prior studies, which often isolate performance under simplified flow regimes, the present work conducts experimental testing under variable flow velocities, nanoparticle volume fractions, and realistic temperature differentials. These test conditions emulate the actual operating range of vehicle engines. Moreover, this research incorporates a multilayer feedforward ANN model trained using a substantial dataset collected from experimental trials. The model includes nanoparticle type, concentration, flow rate, and temperature difference as inputs and is developed as a robust surrogate tool for predicting performance behavior across a wide operational spectrum.

What distinguishes this study is its integrated focus on both thermal and hydraulic performance metrics, coupled with the application of an artificial neural network (ANN) model trained on experimentally validated nanofluid data. By incorporating key indicators such as engine block temperature, coolant outlet temperature, and pressure drop, the study provides a comprehensive, multi-domain assessment of cooling system behavior. The ANN model further enhances the analysis by enabling rapid sensitivity assessments and optimization across different nanoparticle types and concentrations. Although previous studies have individually examined aspects of nanofluid-based cooling performance, this investigation, based on the authors’ knowledge, is the first to offer a comparative, ANN-supported evaluation of Al₂O₃, TiO₂, and CuO nanofluids under automotive-relevant thermal and hydraulic conditions. The outcomes have broader relevance not only for internal combustion engine (ICE) vehicles but also for electric and hybrid vehicles, which require advanced cooling solutions for battery systems, power electronics, and electric motors. This research addresses existing knowledge gaps by experimentally assessing the performance of selected nanofluids in an automotive cooling system and validating the results through an ANN-based surrogate model trained on real-world operational data.

2. Methodology

The methods and modeling approaches utilized to examine the thermal and hydraulic performance of AlO₃, TiO₂, and CuO nanofluids (with volume concentrations varying from 0.5 percent to 2.0 percent) for cooling automotive engines are presented in depth in this section. The approach combines data collection and error analysis, governing physics and heat transfer modeling, experimental setup, nanofluid synthesis and characterization, and artificial neural network-based prediction modeling (ANN).

2.1. Experimental Setup and Nanofluid Preparation

A closed-loop experimental test rig was developed to emulate the thermal-hydraulic environment of an automotive engine cooling system (Figure 1A,B). The system comprised a copper tube test section with a length of 1.0 m, an inner diameter of 12 mm, and a wall thickness of 1 mm, serving as a surrogate engine coolant channel. Thermal loading was provided by an electric heater with a maximum capacity of 2 kW, while a variable-speed centrifugal pump allowed precise flow rate control between 1 and 5 L/min. The setup was insulated with 20 mm thick fiberglass to minimize heat losses

High-accuracy instrumentation included Type-K thermocouples (±0.5 °C, Omega Engineering, Norwalk, CT, USA, Model KMQSS-125U) placed at the inlet, outlet, and wall at the axial mid-length. Volumetric flow was measured using a rotameter (0–10 L/min, ±1.5% full scale, calibrated against a gravimetric standard), and pressure drops were monitored using a piezoresistive differential pressure transducer (0–1 bar, ±0.25% full scale, Honeywell, Model 26PC). Data acquisition was performed via a National Instruments DAQ system (NI cDAQ-9178, sampling rate 10 Hz) integrated with LabVIEW software (Version 2025 Q3). Calibration of thermocouples was performed against an ice-water bath and a reference platinum thermometer (±0.1 K). Steady-state operation was defined as the condition where all monitored parameters remained stable within ±0.2 °C for at least 10 min before data collection.

Nanofluids were prepared by dispersing high-purity Al₂O₃, TiO₂, and CuO nanoparticles (20–50 nm) into deionized water at volume concentrations ranging from 0.5% to 2.0%. Each sample was magnetically stirred for 30 min and then subjected to ultrasonication at 40 kHz for 2 h to ensure uniform dispersion and minimize agglomeration. The thermophysical properties, namely, density, specific heat, viscosity, and thermal conductivity, were estimated using standard mixture models reported in the literature.

The structural morphology of the nanoparticles used in the study was confirmed using transmission electron microscopy (TEM). As shown in Figure 2, the Al₂O₃ nanoparticles exhibited irregular but mostly spherical shapes with slight clustering. In contrast, Figure 3 highlights the CuO nanoparticles, which were rod-shaped and aligned in bundles, potentially impacting flow resistance and anisotropic heat transfer. The TiO₂ nanoparticles, which are shown in Figure 4, showed a more consistent and evenly distributed spherical shape. These morphological characteristics significantly influenced the thermal and hydraulic behavior of the respective nanofluids during the experiments.

$${\rho }_{nf}=(1-\varphi ){\rho }_f+\varphi {\rho }_p$$

(1)

$${(\rho C_p)}_{nf}=(1-\varphi ){(\rho C_p)}_f+\varphi {(\rho C_p)}_p$$

(2)

$$k_{nf}=k_f\left[{\displaystyle \frac{k_p+2k_f-2\varphi (k_f-k_p)}{k_p+2k_f+\varphi (k_f-k_p)}}\right]$$

(3)

$${\mu }_{nf}={\mu }_f(1+2.5\varphi )$$

(4)

Nanoparticles were dispersed in distilled water without the use of surfactants. The nanofluid suspensions were sonicated using a probe sonicator at 200 W for 60 min to ensure uniform dispersion. The pH of the suspensions was monitored and maintained at approximately 7. All samples were degassed under vacuum for 30 min before filling the test loop to eliminate entrapped air. No filtration was performed. The stability of the nanofluids was assessed using ζ-potential measurements (Malvern Zetasizer Nano ZS, aqueous medium, pH +7, 25 °C) and dynamic light scattering (DLS) to determine particle size distribution, conducted both before and after experimental runs. Suspensions with ζ-potential magnitudes greater than ±30 mV were considered stable, in accordance with standard colloidal stability criteria reported in the literature.

2.2. Experimental Procedure, Data Reduction, and Error Analysis

The test section consisted of a straight copper tube with a length L = 1.0 m, inner diameter D = 12 mm, wall thickness = 1.0 mm, and material roughness Ra ≈ 0.2 μm. The hydraulic diameter Dh was taken as the inner diameter. All thermohydraulic parameters (Re, Pr, Nu, f) were calculated using these dimensions. The tube was insulated with 20 mm thick fiberglass insulation to minimize external heat losses. A closed-loop engine cooling test rig was custom-built to simulate the thermal-hydraulic conditions of an automotive cooling system. The rig consisted of the copper tube test section representing the engine coolant channel, an electric heater with a maximum capacity of 2 kW to provide thermal loading, and a variable-speed centrifugal pump enabling flow rates between 1 and 5 L/min. The system was instrumented with Type-K thermocouples (±0.5 °C) for temperature measurements, a calibrated rotameter (±2%) for flow rate monitoring, and a piezoresistive pressure sensor (±1%) to evaluate pressure drops. Data were acquired using a National Instruments DAQ system integrated with LabVIEW software, and steady-state conditions were confirmed by ensuring parameter stability for at least 10 min before measurements.

Nanofluids were pumped through a test section in this investigation at different flow rates, ranging from 1 to 5 L per minute. The heater power input was carefully adjusted between 500 and 2000 watts to simulate actual engine thermal loads. A thorough set of measurements, including the input and outlet fluid temperatures, the test section’s wall surface temperature, the volumetric flow rate, and the associated pressure drop across the section, was recorded after steady-state conditions were reached.

The heat transfer rate was determined using the measured temperature difference and flow rate data. Based on the calculated heat transfer rate and the temperature gradient between the fluid and the wall surface, the convective heat transfer coefficient was derived to evaluate the thermal performance of the nanofluids. In addition, the Darcy friction factor was computed to quantify the hydraulic losses, which depend on the pressure drop, fluid velocity, test section length, and hydraulic diameter. These calculated parameters were essential for assessing the combined thermal and hydraulic performance of the nanofluids under simulated engine cooling conditions.

$$Q=\dot{m}{C}_{p,nf}(T_{out}-T_{in})$$

(5)

$$\dot{m}={\rho }_{nf}\dot{V}$$

(6)

$$h={\displaystyle \frac{Q}{A(T_w-T_b)}}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\mathrm{where}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}\hspace{0.33em}{T}_b={\displaystyle \frac{T_{in}+T_{out}}{2}}$$

(7)

$$f={\displaystyle \frac{2\Delta P{D}_h}{{\rho }_{nf}L{v}^2}}$$

(8)

The test loop was operated as a closed circuit under atmospheric pressure, with distilled water used as the base fluid unless otherwise specified. All experimentally measured bulk coolant temperatures remained below 100 °C. Any reported values exceeding 100 °C in the figures are derived from theoretical extrapolations using the ANN model or empirical correlations rather than direct measurements. No pressurization hardware was employed; therefore, the boiling point under 1 atm was limited to 100 °C. Results above this threshold should be interpreted as computational predictions rather than experimental data. Future work will incorporate pressurized loops and water–ethylene glycol mixtures to extend the safe operating temperature range.

The volumetric flow rate (Q_v, L/min) was measured using a calibrated flowmeter and converted to average velocity via u = Q_v/A, where A = π(D/2)² is the tube cross-sectional area. For consistency between experimental data and modeling, the ANN framework was trained using velocity as the input parameter rather than raw flow rate.

2.3. Governing Equations and Heat Transfer Modeling

The modeling of fluid flow and heat transfer in the nanofluid-based cooling system was based on fundamental conservation equations. These included the continuity equation, which ensures mass conservation in the flow domain; the Navier–Stokes equations that govern the momentum conservation under viscous flow conditions; and the energy conservation equation to account for heat exchange between the fluid and its surroundings.

To simulate the engine’s thermal load, a heat generation model was employed, relating the thermal power input to the coolant mass flow rate value of the coolant and the cooling system’s thermal efficiency. In the experiments, the heat input was supplied by an electrical heater with a maximum capacity of 2 kW. The actual power delivered was calculated directly as Q = U⋅I (voltage × current) and monitored using a calibrated digital power meter. Any references the value pertain solely to the theoretical engine-model scenario and are not part of the present experimental setup. This approach enabled the quantification of the heat that must be dissipated by the cooling system under varying engine loads. The convective heat transfer performance of the nanofluids was assessed by comparing the experimentally derived heat transfer coefficients with those predicted by classical correlations. Specifically, the Dittus–Boelter correlation was used, which requires the computation of the Reynolds number and the Prandtl number to estimate the Nusselt number. The convective heat transfer coefficient was then deduced from the Nusselt number, hydraulic diameter, and thermal conductivity of the fluid.

To account for potential phase change effects under high thermal loads, boiling heat transfer was theoretically considered using the Rohsenow correlation. This empirical model relates the surface heat flux to the temperature difference between the heated wall and the fluid’s saturation temperature, incorporating fluid-specific boiling coefficients and surface properties. While these equations provided a theoretical framework for comparison and validation of the experimental results, it is important to note that the experiments were conducted entirely under single-phase conditions, and no boiling occurred.

$$\nabla \cdot v=0$$

(9)

$${\rho }_{nf}(v\cdot \nabla )v=-\nabla p+{\mu }_{nf}{\nabla }^{2}v$$

(10)

$${\rho }_{nf}{C}_{p,nf}(v\cdot \nabla T)=k_{nf}{\nabla }^{2}T$$

(11)

$$Q={\eta }_{cool}{\dot{m}}_{fuel}LHV$$

(12)

$$Re={\displaystyle \frac{{\rho }_{nf}v{D}_h}{{\mu }_{nf}}}$$

(13)

$$Pr={\displaystyle \frac{{\mu }_{nf}{C}_{p,nf}}{k_{nf}}}$$

(14)

$$Nu=0.023R{e}^{0.8}P{r}^{0.4}$$

(15)

$$h={\displaystyle \frac{Nu\cdot k_{nf}}{D_h}}$$

(16)

$$q^{″}={\mu }_f{h}_{fg}{\left({\displaystyle \frac{g({\rho }_l-{\rho }_v)}{\sigma }}\right)}^{0.5}\left({\displaystyle \frac{C_{sf}}{P{r}^n}}\right){(T_w-T_{sat})}^3$$

(17)

$$U_R=\sqrt{{\displaystyle \sum _i{\left(\frac{\partial R}{\partial x_i}{U}_{x_i}\right)}^2}}$$

(18)

2.4. Artificial Neural Network (ANN) Development

The thermal and hydraulic behavior of nanofluids in automotive engine cooling applications was predicted using feedforward artificial neural networks (ANNs). The dataset comprised 540 experimentally measured data points for Al₂O₃, TiO₂, and CuO nanofluids, which were divided into training (70%, 378 samples), validation (15%, 81 samples), and testing (15%, 81 samples) subsets to ensure balanced distribution and robust model generalization.

The input features included nanoparticle type (one-hot encoded), volume fraction (ϕ = 0.5–2.0%), flow velocity (v = 1.5–2.5 m·s⁻¹), inlet temperature (Tin), and temperature difference (ΔT). The ANN was trained separately for each target output: convective heat transfer coefficient (h, W·m⁻²·K⁻¹) and Darcy friction factor (f, dimensionless). Categorical encoding of nanoparticle type (Al₂O₃, TiO₂, CuO) ensured equal weighting for each category, avoiding ordinal bias (Figure 5).

The network featured two hidden layers with 10 and 6 neurons, respectively, employing the rectified linear unit (ReLU) activation function to capture nonlinear relationships effectively. A single linear output node was designated for predicting the target variables, h and f (Figure 6). Initial weights and biases were set using the Nguyen–Widrow method to improve stability and avoid convergence to local minima. Training was performed in MATLAB R2023b (MathWorks, Natick, MA, USA) using the Adam optimizer (learning rate = 0.001), with a batch size of 32 over 500 epochs, incorporating early stopping with a patience of 50 epochs. Regularization strategies included L2 weight decay (0.001) and dropout (0.2).

Model validation employed 5-fold cross-validation. Performance metrics, including mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²), are evaluated.

2.5. Experimental Accuracy and Data Reliability

To ensure the reliability and reproducibility of the experimental results, a comprehensive uncertainty analysis was performed. High-precision sensors were employed throughout the setup, including calibrated K-type thermocouples with an accuracy of ±0.5 °C for temperature measurement, a rotameter for volumetric flow rate with a range of 0–10 L/min and ±1.5% full-scale accuracy (calibrated against a reference standard), and a differential pressure transducer with a range of 0–1 bar and ±0.25% full-scale accuracy (calibrated before testing). These instruments were selected to minimize measurement errors and maximize confidence in the collected data.

Uncertainty propagation was carried out using standard error analysis, with combined uncertainties of measured parameters analytically evaluated and applied to the heat transfer and pressure drop calculations. The convective heat transfer coefficient and Darcy friction factor showed uncertainties of ±4.8% and ±5.9%, respectively, which are within acceptable limits for thermal-fluid experiments. Nanofluid stability was confirmed through visual inspection, which revealed no sedimentation or agglomeration, and zeta potential measurements above ±30 mV, indicating good dispersion and electrostatic stability. This validated the assumption of homogeneous single-phase flow during testing. However, uncertainties in thermophysical property estimation (thermal conductivity, viscosity, density, and heat capacity) were not included; sensitivity analysis varying these inputs by ±5% indicated up to ±4% variation in predicted heat transfer coefficients, highlighting the need for more precise property data in future work.

The reliability of the experimental dataset was supported by quantified uncertainties, consistent testing conditions, and high-quality instrumentation, ensuring that the data used for ANN training were both robust and representative, thereby strengthening their predictive capability for automotive cooling applications. To isolate the effects of nanoparticle concentration and velocity, all measurements were performed under steady-state thermal and hydraulic conditions; however, real cooling systems experience transient phenomena such as temperature cycling and variable loads that were not captured in this study. Consequently, the reported results should be interpreted as baseline steady-state performance, with future work aimed at extending the facility to evaluate transient engine-relevant dynamics for improved practical relevance.

3. Results and Discussion

This section presents a comprehensive analysis that integrates both experimental observations and ANN-based predictive outcomes, with a particular focus on the thermal and hydraulic behavior of Al₂O₃, TiO₂, and CuO nanofluids under engine-relevant operating conditions. The discussion highlights the influence of nanoparticle type, volume fraction, flow velocity, and thermal load on critical performance indicators, namely the convective heat transfer coefficient and pressure drop. Results are systematically organized to reveal performance trends, identify optimal operating regimes, and delineate potential trade-offs, thereby offering practical guidelines for the formulation of nanofluid coolants in advanced thermal management applications. Furthermore, the predictive performance of the developed multilayer feedforward artificial neural network is validated against the experimental dataset, demonstrating an excellent correlation and underscoring its value as a surrogate modeling framework for the rapid design and optimization of nanofluid-based cooling systems.

3.1. Influence of Nanoparticle Type, Volume Concentration, and Flow Conditions on Convective Heat Transfer Performance

Figure 7 presents a comprehensive analysis of how nanoparticle type and volume concentration (ϕ) influence the convective heat transfer coefficient (h) in nanofluids composed of Al₂O₃, CuO, and TiO₂ under varying temperature differences (ΔT), with flow velocity held constant at 1.5 m·s⁻¹ and bulk coolant temperature at 65 °C, conditions representative of engine cooling environments. Across all nanofluid types, a nonlinear rise in h is observed with increasing ΔT, culminating in a maximum followed by a plateau or mild decline. This behavior is indicative of enhanced micro convection due to Brownian motion and thermal boundary layer disturbance at moderate ΔT, which subsequently becomes limited by factors such as nanoparticle agglomeration, viscosity increase, or saturation in thermal conductivity at higher ΔT levels [25,26]. These mechanisms are particularly relevant under high thermal loads encountered in internal combustion engines.

CuO-based suspensions outperform the other two nanofluids in terms of convective heat transfer, with a peak h value of roughly 1.5 × 10⁵ W·m⁻²·K⁻¹ at ϕ = 1.0% and ΔT = 30 °C. This enhancement is attributed to the high thermal conductivity and density of CuO nanoparticles, which promote rapid energy diffusion within the coolant, thereby supporting efficient heat dissipation and mitigating thermal fatigue in engine components. Al₂O₃ nanofluids, although slightly less effective than CuO, reach respectable h values around 1.4 × 10⁵ W·m⁻²·K⁻¹ within the optimal concentration range of ϕ = 1.0–1.5%. Beyond this threshold, thermal gains diminish, likely due to increased fluid resistance and the onset of thermophysical instabilities. TiO₂ nanofluids, while offering chemical stability and wide applicability in passive systems, fall short in high-demand engine-like conditions, especially at elevated concentrations, where their convective performance is notably inferior. These results collectively underscore the significance of nanoparticle thermal properties in dictating nanofluid effectiveness under dynamic thermal environments.

The impact of flow velocity on nanofluid performance is further examined in Figure 8 and Figure 9, which show the convective heat transfer behavior for speeds of 1.5 m·s⁻¹ and 2.5 m·s⁻¹, respectively, at a bulk coolant temperature of 75 °C. For all nanofluids, h rises with ΔT at the lower velocity (Figure 5), peaks, and then slightly declines. With h values of roughly 1.4 × 10⁴ W·m⁻²·K⁻¹ at ϕ = 1.0 percent to 1.5 percent, CuO nanofluids once again show the largest enhancement, followed by Al₂O₃ and TiO₂. The hierarchy of performance is the same, with CuO taking advantage of its better thermal transport capabilities, whereas TiO₂ has higher viscosity and particle clustering effects that reduce convective enhancement, particularly at high concentrations.

All types of nanofluids show a significant increase in h when the flow velocity is increased to 2.5 m·s⁻¹ (Figure 6), underscoring the crucial role that forced convection plays in accelerating heat transfer. At ideal loadings, CuO nanofluids achieve peak values close to 2.0 × 10⁴ W/m²·K, thanks to enhanced flow turbulence and a smaller thermal boundary layer that improves nanoparticle energy transmission. With h values that closely resemble those of CuO, AlO₃ nanofluids also perform well under these circumstances, confirming their potential in high-flow-rate systems where colloidal and thermal stability are crucial. On the other hand, TiO₂ nanofluids still perform poorly, indicating that even with higher convective forces, their thermophysical constraints still exist. The comparison between Figure 5 and Figure 6 clearly illustrates the amplifying effect of velocity on nanofluid heat transfer, particularly when coupled with high-conductivity particles such as CuO.

The convective heat transfer properties at a fixed flow velocity of 2.5 m·s⁻¹ and a coolant temperature of 85 °C are examined in greater detail in Figure 10, which also evaluates a broader concentration range (ϕ = 0.0–2.0 percent) for nanofluids made with AlO₃, Cu, and TiO₂ nanoparticles. In all fluids, h shows a sharp rise with ΔT until around 25–30 °C, at which point it plateaus or declines, which is in keeping with previous data. According to this, there may be a first phase that is dominated by boundary layer enhancement and thermal diffusion caused by nanoparticles, followed by a decline in returns due to rheological constraints, nanoparticle clustering, or the start of thermal transport saturation.

CuO-based nanofluids exhibit the highest thermal performance, reaching a peak heat transfer coefficient of approximately 1.78 × 10⁴ W·m⁻²·K⁻¹ at a nanoparticle volume fraction of 1.0%. This superior performance is attributed to copper’s exceptionally high thermal conductivity (40 W/m·K) and excellent thermal diffusivity. Among the nanofluids, CuO outperforms others due to its higher thermal conductivity (33–40 W/m·K) compared to Al₂O₃ (25–30 W/m·K) and TiO₂ (8–12 W/m·K). But when this concentration is exceeded, h slightly decreases, most likely as a result of reduced nanoparticle mobility and turbulence suppression brought on by viscosity. Following closely after are AlO₃ nanofluids, which peak at ϕ = 1.0–1.5 percent with h values close to 1.65 × 10⁴ W/m²·K. AlO₃ is a good option for moderate-to-high-performance applications due to its moderate density and stability in suspension, despite having a lower thermal conductivity than Cu. The least favorable thermal response is once again displayed by TiO₂ nanofluids, especially at ϕ = 2.0 percent, where the maximum h stays below 1.55 × 10⁴ W/m²·K. Beyond 25 °C, their relatively flat h–ΔT curves indicate little advantage from higher thermal gradients, most likely as a result of low effective thermal conductivity and premature agglomeration.

Overall, the findings establish a clear performance hierarchy among the tested nanofluids under engine-relevant flow and temperature conditions. CuO nanofluids consistently delivered the highest convective enhancement due to their superior thermophysical properties and strong response to forced convection, though their practical adoption is constrained by higher material costs, increased viscosity at elevated concentrations (raising pumping power), and potential risks of erosion or corrosion in cooling components. Al₂O₃-based fluids provided a balanced compromise between heat transfer performance and stability, particularly at moderate concentrations, while TiO₂ nanofluids, despite their chemical stability and environmental safety, exhibited limited suitability for high-flux applications and are better positioned for low-demand or passive systems. To advance toward real-world implementation, future efforts must optimize nanoparticle loading to balance heat transfer gains against pressure drop penalties and conduct cost–benefit assessments that account for durability, system compatibility, and overall efficiency improvements.

The optimal nanoparticle concentration range of ϕ = 0.5–1.5% for all tested nanofluids is particularly important from the perspective of engineering design. Beyond this range, thermal performance tends to plateau or diminish, primarily due to increased viscosity-induced flow resistance and reduced nanoparticle mobility. In the context of automotive thermal management systems, which require compact, high-efficiency designs and stable fluid behavior under transient thermal loads, CuO-based nanofluids demonstrate the most favorable performance characteristics. Their ability to maintain elevated convective heat transfer coefficients (h) across a range of temperature differences (ΔT) and flow velocities indicates strong potential for applications such as radiator downsizing, enhanced heat exchanger efficiency, and improved thermal regulation in both internal combustion and hybrid electric powertrains. Continued advancements in nanofluid formulation and stabilization, particularly at higher concentrations, may further broaden their applicability in next-generation thermal management solutions.

Notably, CuO nanofluids outperformed Al₂O₃ and TiO₂ by approximately 7% and 15%, respectively, in peak convective heat transfer performance at ΔT ≈ 30 °C and ϕ = 1.0%.

3.2. Effect of Nanoparticle Type and Concentration on Engine Block and Coolant Temperature Across Engine RPM

Einstein’s viscosity model, μnf = μf (1 + 2.5ϕ), was employed as a baseline for estimating nanofluid viscosity; however, it is formally valid only for dilute suspensions of spherical particles. Since CuO nanoparticles are rod-like, the Krieger–Dougherty model incorporating a shape factor was also applied to assess sensitivity. Similarly, the Maxwell and Hamilton–Crosser correlations were used to estimate thermal conductivity, with shape factor adjustments (n = for spherical particles, n > 3 for rod-like particles). Experimental determination of nanofluid properties is recommended for future studies to enhance accuracy.

The thermal performance of nanofluids containing Al₂O₃, CuO, and TiO₂ nanoparticles was evaluated over a simulated engine speed range of 1000–6000 rpm, focusing on their effects on engine block and coolant temperatures. Experiments were conducted under controlled conditions of 25 °C ambient temperature, 1 atm pressure, and a nominal airflow velocity of 33.33 m·s⁻¹. Coolant temperature histories were recorded under heat loads corresponding to approximate engine-equivalent power dissipation. It should be emphasized that both the engine speed range and the airflow velocity represent modeled boundary conditions for the test facility and were not physically applied in the experimental loop. The cited airflow values, such as 33.33 m·s⁻¹, pertain solely to theoretical radiator simulations, as no fan or wind tunnel system was incorporated into the actual test rig.

Figure 11 illustrates the variation in engine block temperature across nanoparticle volume fractions (ϕ) from 0.0% to 2.0%, while Figure 12 presents the corresponding coolant temperature profiles. Engine block temperature decreased significantly with increasing RPM for all nanofluids, primarily due to enhanced convective heat transfer from increased turbulence and coolant flow velocity (Figure 8). CuO-based nanofluids demonstrated the greatest cooling effect, especially at ϕ = 1.0–1.5%, reducing temperatures by up to 40–50 °C compared to the base fluid at low RPMs (1000–2500). CuO’s strong inherent thermal conductivity (~33–40 W/m·K), which promotes effective heat transmission, is responsible for this performance. TiO₂ nanofluids showed the least improvement, perhaps because of their lower thermal conductivity (8.4 W/m·K) and agglomeration at higher concentrations, whereas Al₂O₃ nanofluids demonstrated moderate cooling because of their strong stability and thermal conductivity. After ϕ = 1.5 percent, all nanofluids showed a saturation effect, with viscosity rises limiting additional advances.

The coolant temperature profiles are shown in Figure 9, where CuO nanofluids at higher RPMs and ideal concentrations kept the temperature below 140 °C. While TiO₂ demonstrated decreased efficacy at higher loadings because of viscosity and sedimentation effects, Al₂O₃ nanofluids operated consistently in the 145–150 °C range. At low to mid RPMs (1000–3000), the biggest improvements above the base fluid were observed, underscoring the benefit of nanofluids in situations where traditional coolants generally perform poorly. The higher reported values (up to 140–150 °C) were obtained from theoretical extrapolations and ANN forecasts rather than actual observations, and it is significant to highlight that all experimentally recorded bulk coolant temperatures stayed below 100 °C at 1 atm.

The design of vehicle cooling systems is significantly impacted by these discoveries. CuO nanofluids hold particular promise for thermally demanding situations like rapid acceleration and idling, which could lead to reduced radiator size and increased heat exchanger efficiency. Although TiO₂ nanofluids have advantages for the environment, they are less appropriate for high-performance applications than Al₂O₃ nanofluids, which provide a balanced solution combining performance, affordability, and stability.

In summary, nanoparticle selection and concentration optimization are critical for maximizing nanofluid cooling efficiency. The optimal range of ϕ = 1.0–1.5% emerges as a key design parameter for automotive thermal management, with significant potential to enhance thermal regulation in internal combustion, hybrid, and electric vehicle platforms.

At low RPMs (1000–2500), CuO nanofluids reduced engine block temperature by up to 40–50 °C and coolant temperature by approximately 15 °C compared to the base fluid. The trends observed in Figure 8 and Figure 9 strongly support the conclusion that CuO nanofluids are the most effective for high-load cooling applications, particularly at engine speeds below 4500 RPM.

Long-term operation in real cooling systems may pose further difficulties, even while short-term stability evaluations (visual inspection and zeta potential >30 mV) verify stable suspensions during the measurement period. Possible problems include settling in low-velocity areas, corrosion or erosion of pump/radiator materials, interactions with coolant additives, and progressive agglomeration under prolonged high temperatures. Long-term aging tests, surface functionalization, inline filtering, and tailored surfactants are examples of mitigation techniques. For this reason, thorough compatibility and durability research is necessary before field deployment.

3.3. Validation of ANN Model Predictions Against Experimental Data

The dependability and prediction power of the constructed artificial neural network (ANN) model were tested using experimental data from convective heat transfer measurements of nanofluids under engine-relevant thermal conditions. Input parameters were nanoparticle type (Al₂O₃, TiO₂, CuO), volume concentration (ϕ = 0.5 percent –2.0 percent), temperature differential (ΔT), and flow velocity (1.5–2.5 m·s⁻¹), with the convective heat transfer coefficient (h) being the goal output. To guarantee the robustness of the model, experimental measurements were conducted under a wide range of operational situations.

The efficiency of the ANN in capturing the nonlinear interactions between inputs and heat transfer performance is demonstrated by Figure 13, which shows that projected values (hpred) closely match experimental observations (hexp) along the 1:1 reference line. Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and coefficient of determination (R²) were the four statistical measures used to measure the correctness of the model. The ANN demonstrated great generalization and high prediction accuracy with an R² value of above 0.98, low RMSE and MAE values, and an MAPE below 5 percent.

These results validate the ANN model as a reliable surrogate for rapid prototyping and efficient optimization in automotive thermal management systems employing nanofluids.

$$\mathrm{MAPE}={\displaystyle \frac{100}{n}}{\displaystyle \sum _{i=1}^n\left|\frac{h_{\exp ,i}-h_{\mathrm{pred},i}}{h_{\exp ,i}}\right|}$$

(19)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n{\left(h_{\exp ,i}-h_{\mathrm{pred},i}\right)}^2}}$$

(20)

$$\mathrm{MAE}={\displaystyle \frac{1}{n}}{\displaystyle \sum _{i=1}^n\left|h_{\exp ,i}-h_{\mathrm{pred},i}\right|}$$

(21)

$$R^2=1-{\displaystyle \frac{{\displaystyle \sum _{i=1}^n{\left(h_{\exp ,i}-h_{\mathrm{pred},i}\right)}^2}}{{\displaystyle \sum _{i=1}^n{\left(h_{\exp ,i}-{\overline{h}}_{\exp }\right)}^2}}}$$

(22)

The ANN model produced a high R2 value of 0.985, an MAPE of 3.92 percent, an RMSE of 4.56 × 10³ W/m²⋅K, and an MAE of 3.11 × 10³ W/m²⋅K. With the ANN accounting for roughly 98.5% of the variation in the experimental dataset, these results show outstanding prediction accuracy. The model’s dependability for engineering applications is further supported by the MAPE value, which is much below the generally acknowledged 5 percent criterion.

These findings are consistent with prior studies in the field. Yeneneh et al. (2025) [12] reported R² values ranging from 0.96 to 0.99 for Al₂O₃ and CuO nanofluids using ANN-based regression under turbulent flow conditions. Bandarra Filho et al. (2025) [25] found MAPE values as low as 4.2% when validating ANN-predicted Nusselt numbers for TiO₂-water nanofluids in radiator systems. Moreover, the enhancement trend observed for CuO nanofluids in this study aligns well with thermal behavior profiles reported in automotive cooling applications, supporting the model’s generalization.

In terms of nanoparticle-specific performance, CuO nanofluids showed the most consistent prediction accuracy with <3% deviation across all ΔT and ϕ levels. Al₂O₃ nanofluids exhibited moderate deviation at higher concentrations (ϕ > 1.5%), potentially due to increased viscosity effects not fully modeled. TiO₂ nanofluids displayed the largest discrepancies at low flow velocities and high-volume fractions, where particle clustering and resistance effects introduced nonlinearities beyond the current ANN’s resolution.

In conclusion, the ANN model demonstrated high generalization capacity and accuracy in predicting the heat transfer behavior of different nanofluids. This validation confirms the suitability of neural network-based models for use as surrogate predictive tools in thermal management design, reducing the dependence on extensive experimental campaigns while facilitating faster system-level optimization.

These findings are consistent with previous studies reporting R2R^2R2 values between 0.96 and 0.99 for Al₂O₃ and CuO nanofluids using ANN-based regression under turbulent flow conditions [12] and mean absolute percentage error (MAPE) values as low as 4.2% for TiO₂ water nanofluids in radiator systems. The trends observed for CuO nanofluid performance also align with the reported thermal behavior in automotive cooling applications, supporting the generalization capability of the ANN model. Residual analyses for the model predictions are shown in Figure 14 and Figure 15. Figure 14 presents the residuals of the convective heat transfer coefficient (h) against experimental values for Al₂O₃, TiO₂, and CuO nanofluids, while Figure 15 shows the residuals of the Darcy friction factor (f) relative to experimental measurements. These plots confirm minimal bias and tight clustering around zero, indicating high predictive accuracy.

In summary, the ANN models demonstrated excellent performance and strong generalization in predicting both thermal and hydraulic behavior across diverse nanofluids. Validation results confirm the suitability of the models as surrogate predictive tools for rapid system-level optimization in automotive thermal management, reducing reliance on extensive experimental campaigns and facilitating efficient design and prototyping.

The ANNs were trained using experimental data within a limited domain of nanoparticle concentrations (0.5–2.0 vol.%) and flow velocities (1.5–2.5 m·s⁻¹). Within this range, the models achieved excellent accuracy (R2 > 0.98). Predictive capability outside this domain is not guaranteed, and extrapolations to other concentrations, base fluids, or operating conditions should be approached with caution. For practical applications, the ANNs can serve as efficient interpolators within the validated domain, with additional data or retraining required for extension to new regimes.

Two separate ANN models were implemented: one for predicting the convective heat transfer coefficient (h) and another for the friction factor (f). Both models used the same input parameters but were trained and validated independently to ensure optimal performance for each target variable.

3.4. Experimental Uncertainty

To evaluate the correctness of the obtained thermal-hydraulic parameters and the dependability of the experimental observations, a thorough uncertainty analysis was conducted. Throughout the test setup, high-precision sensors were utilized, and the overall uncertainty of computed values, such as the Darcy friction factor and convective heat transfer coefficient, was estimated using standard uncertainty propagation techniques.

Table 1. Summary of instrument accuracy and parameter uncertainties.

Parameter Measured or Calculated	Instrument Used	Accuracy	Estimated Uncertainty (%)
Fluid Temperature (Inlet/Outlet)	K-type Thermocouple	±0.5 °C	±1.8%
Flow Rate	Rotameter	±1.5% of full scale	±2.1%
Pressure Drop	Piezoresistive Pressure Transducer	±0.25% of full scale	±2.5%
Convective Heat Transfer Coefficient (h)	Derived (based on flow rate, ΔT, etc.)	—	±4.8%
Darcy Friction Factor (f)	Derived (based on pressure drop, flow)	—	±5.9%

summarizes the measurement instruments, their accuracy specifications, and the resulting uncertainty in key performance metrics.

The combined uncertainty of each derived parameter was calculated using the root sum square (RSS) method of uncertainty propagation, expressed mathematically as

$$U_R=\sqrt{{\left({\displaystyle \frac{\partial R}{\partial x_1}}\cdot U_{x_1}\right)}^2+{\left({\displaystyle \frac{\partial R}{\partial x_2}}\cdot U_{x_2}\right)}^2+\cdots +{\left({\displaystyle \frac{\partial R}{\partial x_n}}\cdot U_{x_n}\right)}^2}$$

(23)

The uncertainties associated with the convective heat transfer coefficient and the Darcy friction factor were quantified using standard uncertainty propagation methods, which account for the combined effects of all contributing measurement errors. These calculated margins fall within the acceptable range for experimental thermal-fluid research, thereby supporting the reliability of the results. The uncertainty in the convective heat transfer coefficient (h) determined based on flow rate, fluid temperature difference, and thermophysical properties was specifically evaluated by applying partial derivatives to the governing equation with respect to each input parameter, combined with their respective uncertainties.

Note that the current uncertainty analysis only takes measurement-related errors into consideration (temperature, pressure drop, and flow rate). The estimation uncertainties of thermophysical properties (specific heat, density, viscosity, and thermal conductivity) were not specifically measured. These characteristics may have intrinsic error (usually ±3–10 percent) because they were derived from empirical models and correlations found in the literature. By altering the input properties by ±5 percent and recalculating the convective heat transfer coefficient, a sensitivity analysis was conducted to evaluate the effect of these uncertainties. The relative change in h, according to the results, was within ±4 percent, which is less than the experimental scatter but still substantial in several operating regimes. This emphasizes the significance of enhanced databases of nanofluid properties and in situ measurements for more precise predictions at the system level.

To ensure the consistency and stability of the nanofluids during testing, both visual inspections and zeta potential analyses were conducted. All nanofluid samples exhibited zeta potential values greater than ±30 mV, confirming electrostatic stability, uniform nanoparticle dispersion, and the absence of significant agglomeration throughout the experimental duration. These findings not only validate the single-phase flow assumption but also confirm that the thermophysical properties remained consistent across all experimental runs.

4. Conclusions

A thorough experimental and theoretical evaluation of nanofluids made of AlO₃, TiO₂, and CuO nanoparticles for possible use in car engine cooling systems is provided in this work. CuO-based nanofluids proved to be superior to the other two in terms of the convective heat transfer coefficient through systematic thermal and hydraulic testing under engine-like flow and temperature conditions, especially within an ideal concentration range of 1.0–1.5 percent. While TiO₂ nanofluids showed greater sensitivity to increased viscosity and particle agglomeration and low thermal enhancement, Al₂O₃ nanofluids offered a well-balanced trade-off between stability and performance. Notably, thermal performance benefits were lost or reversed after the ideal loading threshold was reached, highlighting the necessity of striking a balance in practical system design between tolerable pressure drop, fluid rheology, and heat transfer improvement.

A key contribution of this work is the development and validation of a multilayer feedforward artificial neural network (ANN) model trained on a large experimental dataset. The model successfully predicted thermal performance metrics such as the convective heat transfer coefficient and Darcy friction factor, achieving a coefficient of determination (R²) of 0.985 and a mean absolute percentage error (MAPE) below 5%. These results affirm the ANN model’s robustness and its applicability as a surrogate modeling tool for rapid thermal system evaluation and optimization. The incorporation of nanoparticle type, volume concentration, temperature differential, and flow velocity as input features allowed the model to capture complex nonlinear interactions often overlooked by conventional empirical correlations. By leveraging the ANN’s predictive capability, engineers and researchers can reduce their dependence on time-consuming and resource-intensive experimental trials when designing high-efficiency cooling systems.

The results of this study are more broadly applicable to hybrid and electric vehicle platforms, which require sophisticated temperature control for parts like battery packs, electric motors, and power electronics, than just the immediate context of internal combustion engine cooling. A fundamental basis for choosing appropriate coolant formulations suited to various operating situations and thermal loads is provided by the performance hierarchy developed among the investigated nanofluids. CuO nanofluids exhibit great promise for high-load, small-system applications due to their exceptional thermal conductivity. Al₂O₄ nanofluids, on the other hand, provide a more stable and affordable option for modest cooling needs, which makes them especially appealing for systems where reduced viscosity and long-term chemical compatibility are crucial. TiO₂ may still be useful in passive or low-demand systems even though it is not as good for high-performance applications.

Nevertheless, several limitations must be acknowledged. The study primarily focused on steady-state thermal and hydraulic performance and did not address long-term operational stability under cyclic engine loads, chemical interactions with metal components, or nanoparticle dispersion longevity over extended usage periods. Additionally, while the ANN model demonstrated high accuracy within the tested parameter space, its extrapolation to extreme conditions or hybrid nanofluid blends requires further verification. Environmental and economic considerations, such as nanoparticle manufacturing cost, recyclability, and potential health impacts, were also beyond the scope of this investigation and merit further exploration before large-scale adoption in commercial vehicles.

Future studies should focus on developing hybrid nanofluid formulations with enhanced thermal and rheological properties, assessing durability under dynamic engine conditions, and evaluating long-term compatibility with radiator and pump materials. Scaling up to full vehicle-level testing would allow comprehensive assessment of nanofluid effectiveness in reducing radiator size, improving energy efficiency, and extending component life. Concurrently, further development of the ANN framework—through integration with real-time sensor data and optimization algorithms could enable adaptive thermal control in smart vehicle systems. Additional priorities include expanding datasets to improve predictive accuracy, testing under rapid thermal cycling, and quantifying uncertainties in thermophysical parameters.

Although CuO nanofluids demonstrate promising thermal performance, challenges such as cost, erosion, viscosity-induced pumping penalties, and long-term durability remain. Future work should also address transient thermal cycling, quantify long-term stability, explore pressurized or ethylene glycol-based coolants to extend safe operating ranges, and broaden the ANN training domain. Clear specification of property uncertainties, CuO terminology, and ANN outputs will ensure an accurate representation of the study’s scope and limitations.

Finally, by combining rigorous experimental validation with machine learning-based predictive modeling, this study provides a solid basis for the application of oxide-based nanofluids in next-generation engine cooling systems. Compact, effective, and adaptable automobile thermal management solutions are made possible by nanofluids and data-driven methodologies, which have been shown to be able to precisely model and optimize thermal-hydraulic behavior.