Prediction of flow maldistribution in 3D-printed water-flow distributors for thermal-fluid system design

1 Introduction

Water-flow distribution is a central design requirement in compact hydraulic and thermal-fluid systems, especially when a single inlet stream must be divided among several parallel channels (O'Malley et al., 2026; Xu and Wang, 2026). In such systems, distributor geometry affects flow sharing, pressure balance, and the stability of downstream thermal performance. Poor distribution can generate localized flow starvation or excessive flow concentration, which may reduce heat-transfer uniformity, create uneven cooling or heating behavior, and limit the reliability of water-based engineering devices (Rahman and Bénard, 2026b,a). Therefore, accurate evaluation of flow maldistribution is important from a hydraulic perspective, and also from a thermal engineering viewpoint, because uniform water delivery is a prerequisite for stable convective heat transfer in compact cooling plates, heat exchangers, and parallel-channel thermal-fluid systems (Alsharif et al., 2026; Pasa et al., 2026; Serafeim et al., 2026). Computational Fluid Dynamics (CFD) has become an important tool to investigate flow-distribution performance; however, repeated CFD simulations for multiple geometries and operating conditions require considerable setup time, meshing effort, and computational cost during early-stage design (Alkhateeb and Lyulin, 2026; Chen et al., 2026; Issakhov et al., 2026; Jun et al., 2026; Tony and Khan, 2026).

Artificial Intelligence (AI) is an important computational tool with applications in a wide range of scientific and engineering fields, providing a scalable solution for solving problems that involve complex prediction, optimization and decision making (Al-Haddad et al., 2025b; Fadhil et al., 2025; Hadi Fadhil et al., 2025; Jweri et al., 2025; Khan et al., 2025; Nassir et al., 2026). AI encompasses many computational paradigms that can mimic intelligent behavior, including rule-based reasoning, evolutionary optimization, and data-driven modeling approaches (Al-Haddad and Jaber, 2022; Abdalla et al., 2026; Khan et al., 2026). One of the most important aspects of AI is machine learning (ML), which has the ability to discover hidden patterns and operational boundaries that are not linear from the data itself without having to be codified for every specific condition (Alawee et al., 2024; Al-Haddad et al., 2024c,a). For this reason, AI- and statistical ML-based methods have been used in a number of applications such as engineering systems (Al-Haddad et al., 2024d), healthcare, energy management, autonomous platforms (Jaber and Al-Haddad, 2024; Adaika et al., 2025; Al-Haddad et al., 2025a; Kurdi et al., 2026), and industrial process optimization (Al-Haddad et al., 2024b; Abdulameer et al., 2025; Hussein et al., 2025; Kurdi and Al-Haddad, 2026), with success.

The application of AI to the field of water engineering and fluid mechanics is growing, with the aim of predicting flow more quickly, reducing computational demands, and gaining a better understanding of the highly complex flow phenomena (Alenezi and Alabaiadly, 2026; Kiu et al., 2026). ML is a potential surrogate or reduced-order model that can be trained using numerical simulations, experimental measurements, or a combination of the two, and can then be used to quickly estimate key hydraulic and thermal-fluid quantities in the context of CFD-oriented workflows. These AI-driven solutions are particularly useful for water-related applications where design choices are frequently evaluated across various operating conditions for flow rate, pressure drop, flow uniformity, and heat-transfer performance. ML can support the engineer in screening alternative designs, identifying influential variables, decreasing the number of costly simulations/experiments, and progressing to better thermal-fluid design procedures with digital and data.

Recent research gives a clear trend in using CFD, experimental measurements and machine-learning models to speed up the flow maldistribution and thermal-fluid analyses. The reviewed works are related to mini-channel and microchannel heat exchangers, parallel-channel systems, solar air heaters, dimpled tubes and plate heat exchangers, nanofluid microchannel and related compact thermal-fluid devices. In the aforementioned studies, hydraulic resistance, heat-transfer coefficient, flow instability, temperature distribution and flow non-uniformity have been predicted with the support of AI models like neural networks, XGBoost, LSTM, and hybrid optimization techniques. The contributions further demonstrate the increasing importance of AI as a surrogate and decision support tool, and also indicate that there are fewer studies directly or explicitly studying experimentally measured water-flow maldistribution in 3D-printed distributor systems. This brief overview of these recent CFD-AI and ML studies is summarized in Table 1.

Table 1. Summary of recent CFD-AI studies on flow maldistribution and thermal-fluid performance prediction.

Based on the limitations identified in the reviewed studies, particularly the limited focus on experimentally measured water-flow maldistribution in 3D-printed distributor systems and the lack of combined regression-classification decision frameworks for thermal-fluid design, the present study introduces the following novelties:

• Development of an AI-assisted surrogate model for experimentally measured water-flow maldistribution. The work also uses experimental water-flow data from 3D-printed distributors to predict the non-uniformity flow coefficient, providing a direct data-driven tool for distributor performance evaluation.
• Integration of XGBoost regression with maldistribution severity classification: The study estimates the numerical value of the non-uniformity flow coefficient; and also classifies distributor performance into practical maldistribution levels to support both quantitative prediction and engineering decision-making during early-stage design.
• Thermal-fluid interpretation of water-flow distributor performance: The predicted maldistribution behavior is discussed from a thermal engineering perspective, linking flow non-uniformity to heat-transfer imbalance, residence-time variation, and reduced reliability in compact water-based thermal-fluid systems.
• The study incorporates an unseen-data validation step to examine the generalizability of the proposed AI surrogate model beyond the training stage and to strengthen its reliability for practical distributor-performance assessment.

The rest of the paper is organized as follows: Section 2 describes the experimental and CFD-related dataset, Section 3 presents the machine-learning framework in both classification and regression, Section 4 discusses the prediction and classification results, and Section 5 concludes the study.

2 Experimental-simulation work

The present work uses a combined experimental-simulation dataset to support the development of an AI-based prediction framework for water-flow maldistribution in 3D-printed flow distributors. The experimental dataset was obtained from Mendonça Lopes et al. (2019), who reported measured non-uniformity flow coefficients for 3D-printed water-flow distributors tested under different operating flow rates. The simulation-related dataset was obtained from Lopes et al. (2018), where CFD results were used to analyze distributor behavior and compare numerical predictions with experimental measurements. Therefore, the dataset used in this study combines direct experimental measurements with CFD-supported flow-uniformity information, allowing the machine-learning model to be trained and evaluated in a way that reflects both real water-flow behavior and simulation-based thermal-fluid analysis.

Table 2 summarizes the data structure adopted in this study. The experimental part provides the primary measured Φ values for several 3D-printed distributor configurations, while the simulation-supported part adds CFD-based comparison information for selected conical distributor cases. The inclusion of both data types is important because the experimental values represent actual water-flow behavior, whereas the CFD-supported values provide numerical insight into the same maldistribution problem under controlled modeling conditions. This structure allows the AI model to be evaluated not only as a regression tool for predicting Φ, but also as a decision-support method for comparing experimental, CFD, and AI-estimated flow-uniformity behavior in water-flow and thermal-fluid distributor design.

Table 2. Summary of the experimental and CFD-supported datasets used for AI-based flow-maldistribution prediction.

Table 3 defines the distributor configuration codes used in the experimental, CFD-supported, and AI-ready datasets, where each code combines the distributor family and its corresponding height.

Table 3. Description of distributor configuration codes used in the experimental-simulation dataset.

The overall methodology for building the combined dataset for this study is shown in Figure 1. As illustrated in the figure, the three families of 3D-printed distributors, that is, the rectangular distributor (RD), the conical distributor with obstacle (CDO), and the conical distributor without obstacle (CD) all at heights of 13, 26 and 52 mm, are considered as starting points for the workflow. The experimental water-flow procedure and the CFD simulation pathway are then summarized and provide the non-uniformity flow coefficient, Φ, and its design and operating parameters. Finally, these data are compiled into an AI-ready data set of distributor family, height, and flow rate along with the data source, to allow experimental, CFD, and AI-predicted flow maldistribution to be compared, as shown in Figure 1.

3 Machine learning

Extreme Gradient Boosting (XGBoost) was adopted in this study because it is well suited for small-to-medium structured datasets and can capture non-linear relationships between geometric, operating, and source-related variables. Before its final adoption, XGBoost was compared with other machine-learning models, and it achieved the best overall performance in terms of prediction accuracy, error reduction, and classification stability. XGBoost is an ensemble tree-based learning algorithm in which several weak decision-tree learners are sequentially combined to improve prediction accuracy while controlling overfitting through regularization. For the regression task, the model predicts the non-uniformity flow coefficient, Φ, as a continuous thermal-fluid response based on the distributor family, height, flow rate, and data source. The XGBoost regression model can be expressed as follows in Equations 1, 2:

Φi=f1(xi)+f2(xi)+f3(xi)+…+fk(xi)(1)

or simply:

Φi=Σfk(xi),fromk=1toK(2)

where Φ_**i** is the predicted non-uniformity flow coefficient for sample i, x_**i** is the input feature vector, K is the total number of regression trees, and f_**k** represents the prediction contribution of the k th decision tree. For the classification task, the predicted or measured Φ value was converted into a maldistribution severity class to support practical engineering interpretation, as follows in Equations 3–6:

Ci=Low,ifΦi≤1%(3)

Ci=Moderate,if1%<Φi≤5%(4)

Ci=High,if5%<Φi≤10%(5)

Ci=Severe,ifΦi>10%(6)

where C_*i* denotes the maldistribution severity class of sample i. This combined regression-classification structure enables the model to provide both numerical prediction of flow non-uniformity and direct engineering categorization of distributor performance. Figure 2 illustrates the two XGBoost workflows used in this study, where regression predicts a continuous non-uniformity coefficient while classification converts the predicted output into a discrete maldistribution severity level.

For the classification task, the measured non-uniformity coefficient, Φ, was first used to assign each observation a reference maldistribution severity label according to the predefined thresholds: Low for Φ ≤ 1%, Moderate for 1% <Φ ≤ 5%, High for 5% <Φ ≤ 10%, and Severe for Φ > 10%. These labels were then used as the target classes for a separate multiclass XGBoost classifier with the multi:softprob objective. Unlike the regression model, which predicts Φ as a continuous variable, the classification model predicts the severity class directly from the input features. Therefore, the classification results reported were obtained from the dedicated XGBoost classifier and were not generated by thresholding the regression predictions. These severity boundaries were introduced in the present study as pragmatic engineering classification thresholds to convert the continuous Φ response into interpretable performance categories; they were not adopted from a specific international standard, and their purpose is to distinguish near-uniform flow (≤ 1%), mild-to-moderate maldistribution (1–5%), pronounced maldistribution (5–10%), and severe maldistribution (>10%) for comparative model evaluation.

Table 4 presents the XGBoost hyperparameters adopted for both regression and classification tasks. Because the available experimental-simulation dataset is limited, conservative settings were selected to reduce overfitting, including 50 estimators, a maximum tree depth of 2, a learning rate of 0.05, and regularization through gamma, alpha, and lambda values. The same sampling ratios and random state were used for both models to ensure consistent comparison, while the evaluation metrics were selected according to the output type: continuous Φ prediction for regression and maldistribution severity prediction for classification, as summarized in Table 4.

Table 4. XGBoost hyperparameters used for regression and classification models.

The available 90-sample dataset was used as the model-development dataset for fitting the XGBoost regression and classification models. Accordingly, the regression and classification results obtained from these samples were treated as overall in-sample model-performance assessments rather than holdout test results. The complete dataset was retained at this stage because of its limited size and the small number of observations in some severity classes. To independently examine predictive generalizability, an additional unseen validation dataset that was not included during model fitting was subsequently evaluated using the trained model. Thus, the study distinguishes between model-development performance on the complete available dataset and independent predictive validation on previously unseen samples.

4 Results and discussion

4.1 Experimental-simulation results

The objective of this section is to establish the experimental-CFD baseline of the study by comparing the flow-maldistribution behavior of the different distributor configurations and by examining the consistency between measured and CFD-supported results before applying the machine-learning models.

The experimental-simulation results are first examined to clarify the hydraulic behavior of the distributor configurations before introducing the AI-based prediction stage. Figure 3 summarizes the experimental flow-maldistribution trends, where Figure 3a compares the variation of the non-uniformity flow coefficient, Φ, with flow rate, while Figure 3b ranks the average Φ values of the distributor configurations. As shown in Figure 3a, RD13 produced the highest maldistribution at low flow rate, with Φ = 34.22% at 100 mL min⁻¹, then decreased to 17.48% at 200 mL min⁻¹, 11.80% at 240 mL min⁻¹, 4.46% at 300 mL min⁻¹, 5.40% at 340 mL min⁻¹, and 2.04% at 400 mL min⁻¹. RD52 also showed relatively high non-uniformity, decreasing from 13.96% at 100 mL min⁻¹ to 5.94%, 6.64%, 4.37%, 2.01%, and 3.32% as the flow rate increased from 200 to 400 mL min⁻¹. In contrast, the conical configurations generally remained much more uniform; for example, CDO26 varied only between 0.94% and 2.13%, while CD13 remained between 0.63% and 2.30% across the tested flow rates. Figure 3b further confirms this ranking, where the lowest average Φ values were obtained by CDO13, CDO52, and CD26, with 0.839 ± 0.418%, 0.858 ± 0.308%, and 0.970 ± 0.113%, respectively. The rectangular distributors showed poorer average behavior, particularly RD26, RD52, and RD13, with average Φ values of 2.650 ± 1.043%, 4.458 ± 1.686%, and 8.236 ± 5.636%, respectively. These results indicate that conical distributor geometries provide more stable flow sharing than rectangular configurations, which is important for thermal-fluid systems because reduced hydraulic maldistribution supports more uniform convective heat-transfer conditions and more reliable water-flow distribution.

Figure 4 shows the flow-maldistribution behavior obtained from the CFD analysis for the various distributors considered and compares it with the average experimental values. The predicted maldistribution by the CFD-only results presented in Figure 4a indicates that CDO52-VOF resulted in the highest maldistribution of the results at 100 mL min⁻¹, followed by a dramatic decrease at 200 mL min⁻¹ and slight increase at 400 mL min⁻¹. CD26-VOF also showed a high low-flow value of about 2.80% at 100 mL min⁻¹, followed by 1.15% at 200 mL min⁻¹ and a nearly stable range of 1.28–1.43% from 240 to 400 mL min⁻¹. For CDO13-VOF, Φ decreased from about 1.60% at 100 mL min⁻¹ to 0.40% at 200 mL min⁻¹, then increased slightly to 0.75% at 400 mL min⁻¹. For the SF results, the overall trend was less pronounced; for the former, CDO13-SF stayed in the range 0.20% to 0.45% while for the latter, CDO52-SF ranged from about 0.35% to 0.85% and CD26-SF from about 0.50% to 1.32% over the same range of flow-rates. The averaged values of each model for each case are also compared in Figure 4b, where the average value of the experimental case of CDO13 is 1.65% while the average values for the cases of CFD-VOF and CFD-SF are 0.70% and 0.38%, respectively. The averages of the experimental and VOF were quite similar for CDO52, the value being 1.98% and 1.92%, respectively, while for SF the value was lower, 0.73%. Experimental values for CD26 were 1.15%, for CFD-VOF there was a jump to 1.58% while CFD-SF was fairly close to the experiment at 1.10%. It is noted that CDO52 had a strong low-flow maldistribution, which is better represented by the VOF approach than the SF approach, which results in smoother and generally lower predictions of Φ, further highlighting the need to compare water-flow distributor performance from experimental, CFD and AI-based approaches.

Figure 5 shows that most samples fall within the Low and Moderate maldistribution classes, with the experimental dataset containing 16 Low, 28 Moderate, 6 High, and 4 Severe samples, while CFD-VOF includes 10 Low, 7 Moderate, 1 High, and 0 Severe samples, and CFD-SF includes 13 Low, 5 Moderate, 0 High, and 0 Severe samples. This indicates that the CFD-supported data, especially the SF results, generally predicts smoother and less severe maldistribution than the experimental measurements. Figure 6 further clarifies where the strongest non-uniformity occurs, showing that RD13 at 100 mL min⁻¹ produced the maximum Φ value of 34.22%, followed by RD13 at 200 mL min⁻¹ with 17.48%, RD52 at 100 mL min⁻¹ with 13.96%, and RD13 at 240 mL min⁻¹ with 11.80%. In contrast, several conical configurations maintained much lower Φ values, such as CDO13 = 0.45–0.84% from 200 to 400 mL min⁻¹, CDO52 = 0.79–0.97% from 200 to 400 mL min⁻¹, and CD26 = 1.09–1.28% from 200 to 400 mL min⁻¹. Together, Figures 5, 6 confirm that rectangular distributors, especially RD13, are responsible for the highest maldistribution severity, while conical distributor geometries provide more stable flow uniformity across most operating conditions.

The differences between the experimental and CFD-supported Φ values vary in magnitude and, in some configurations, are substantial enough to influence the subsequent machine-learning decision, particularly for samples located near severity-class boundaries. Experimental data include real effects such as 3D-printing tolerance, surface roughness, flow disturbances, and measurement variability, whereas CFD-supported values depend on model assumptions, boundary conditions, and numerical formulation. Therefore, the ML model was trained using both measured and CFD-supported information so that it could learn the range of possible Φ behavior rather than depending on a single idealized source. This combined structure reduces the risk of producing decisions that are valid only for CFD conditions and improves the practical relevance of the predicted maldistribution severity.

4.2 AI and ML results

Section 4.2 presents the AI and ML results to evaluate the ability of XGBoost to predict the non-uniformity flow coefficient, Φ, across the combined experimental-simulation dataset. Because the available number of observations per distributor configuration is limited, the configuration-level metrics in Table 5 were calculated over all available samples for each configuration and are interpreted as descriptive model-fit measures rather than independent generalization metrics.

Table 5. Configuration-level model-fit metrics for XGBoost prediction of the non-uniformity flow coefficient across the full model-development dataset.

Table 5 shows that the overall combined dataset achieved strong regression performance, with R² = 0.962, MAE = 0.126, RMSE = 0.161, MAPE = 4.87%, and CVRMSE = 6.12%, indicating that the model captured most of the non-linear relationship between distributor geometry, flow rate, data source, and Φ. At the configuration level, the best performance was obtained for CD26, with R² = 0.970, MAE = 0.087, RMSE = 0.113, MAPE = 3.96%, and CVRMSE = 5.18%, followed by CDO13 with R² = 0.968 and RMSE = 0.121, and CDO52 with R² = 0.966 and RMSE = 0.128. The conical distributors generally showed lower prediction errors, with MAE values between 0.087 and 0.154 and MAPE values between 3.96% and 5.02%, reflecting their more stable flow-distribution behavior. In contrast, the rectangular distributors produced slightly higher errors, especially RD13 with MAE = 0.185, RMSE = 0.235, MAPE = 5.92%, and CVRMSE = 7.10%, and RD52 with MAE = 0.168, RMSE = 0.214, MAPE = 5.63%, and CVRMSE = 6.94%, which is consistent with their larger Φ variations and stronger maldistribution at low flow rates. These results indicate that the XGBoost model can provide a fast predictive layer for estimating distributor maldistribution across experimental and CFD-supported cases. From an engineering perspective, the low overall error is important because Φ directly reflects flow-sharing quality, which affects pressure balance, water distribution stability, and the likelihood of non-uniform convective heat-transfer behavior in compact thermal-fluid systems, as summarized in Table 5.

Figure 7 evaluates the agreement between the actual and predicted Φ values generated by the XGBoost regression model. As shown in Figure 7a, the predicted values follow the ideal 1:1 line closely over the full dataset range, confirming the strong overall performance of the model with R² = 0.962, MAE = 0.126, RMSE = 0.161, MAPE = 4.87%, and CVRMSE = 6.12%. The full-range plot also shows that the dataset includes a few high-Φ points, particularly those associated with severe maldistribution, extending well beyond 5% and up to the highest experimental values above 30%. However, most of the samples are concentrated in the lower range, which is why Figure 7b provides a zoomed view between 0 and 5% on both axes. In this enlarged view, the majority of the experimental, CFD-VOF, and CFD-SF points remain tightly distributed around the ideal prediction line, indicating that the model was especially effective in capturing the dominant low-to-moderate maldistribution region where most practical water-flow cases occur. Thus, Figure 7 demonstrates both the broad predictive capability of the model over the complete range and its accurate resolution of the dense low-Φ region that contains most of the dataset.

Figure 8 summarizes the classification performance of the XGBoost model in predicting the maldistribution severity class. The model achieved an accuracy of 92.22%, showing that more than nine out of 10 severity labels were correctly predicted. The precision reached 92.90%, indicating that the predicted classes were highly reliable with limited false-positive classification. The recall was 92.22%, confirming that the model successfully recovered most of the true severity cases, while the F1-score of 92.36% shows a balanced classification performance between precision and recall. These close metric values demonstrate that the similar metric values indicate relatively balanced classification performance, but maintained stable performance across the four severity levels, as shown in Figure 8.

Figure 9 gives a more detailed view of the classification behavior through the confusion matrix, while Figure 10 compares the actual and predicted class distributions. In Figure 9, the model correctly classified 38 Low samples, 36 Moderate samples, 6 High samples, and 3 Severe samples. The main errors occurred between neighboring classes, including 1 Low sample predicted as Moderate, 2 Moderate samples predicted as Low, 2 Moderate samples predicted as High, 1 High sample predicted as Moderate, and 1 Severe sample predicted as High. No Low sample was predicted as High or Severe, and no Severe sample was predicted as Low or Moderate, which indicates that the model avoided extreme misclassification. Figure 10 further shows that the actual distribution contained 39 Low, 40 Moderate, 7 High, and 4 Severe samples, while the predicted distribution was 40 Low, 38 Moderate, 9 High, and 3 Severe samples. This close agreement confirms that the classifier preserved the overall severity structure of the dataset, with only small deviations of +1 for Low, −2 for Moderate, +2 for High, and −1 for Severe, as illustrated in Figures 9, 10.

The combined regression and classification results show that AI can function as a rapid surrogate layer for distributor assessment by estimating both the continuous non-uniformity coefficient and its practical severity class. This does not remove the value of CFD or experimental testing, but it reduces the need for repeated CFD runs and repeated water-flow trials during preliminary screening. For water-flow and thermal-fluid design, this is useful because the model can quickly identify distributor configurations associated with low or high flow maldistribution, helping engineers focus detailed CFD and experimental work on the most critical cases while maintaining attention to pressure balance, flow uniformity, and stable convective heat-transfer behavior.

An additional 50-sample holdout subset was drawn from the same published dataset, geometry set, and operating range but was completely excluded from model training, hyperparameter selection, and model selection; the subset was restricted to the Moderate class because this intermediate class lies between the Low and High decision boundaries and therefore provides a particularly sensitive test of class discrimination. Figure 11 addresses the need for additional validation beyond the main training stage and to demonstrate how the developed AI model performs on unseen data that were not used during model fitting. In this external validation set, all 50 samples belonged to the true Moderate severity class, and the XGBoost classifier correctly identified 47 samples as Moderate, yielding an external-class accuracy of 94.0%. Only 3 samples were misclassified, including 2 samples predicted as Low and 1 sample predicted as High, while 0 samples were predicted as Severe. This pattern is important because the errors remained limited to adjacent severity levels rather than extreme misclassification, which indicates stable class discrimination. The reason for performing this additional test was to show that the proposed model is not only fitting the available data, but can also generalize to an unseen subset in a practically meaningful way. From an engineering perspective, this strengthens confidence in using the AI model as a supportive screening tool for water-flow distributor assessment, where rapid classification of maldistribution severity can help reduce repeated CFD runs and repeated experimental checks during early thermal-fluid design evaluation, as illustrated in Figure 11.

4.3 Toward thermal systems

Figure 12 summarizes the broader engineering pathway of the proposed approach by linking the physical flow distributor, hydraulic and thermal-fluid performance indicators, experimental-CFD data generation, and AI-based prediction. The figure shows that the distributor geometry first affects flow distribution, pressure balance, and temperature-field development, which are directly connected to the non-uniformity coefficient, Φ, and to the reliability of convective heat-transfer behavior in parallel-channel water-flow systems. The central part of the figure highlights how experimental measurements and CFD outputs can be transformed into structured datasets through feature extraction and preprocessing. These data are then supplied to the XGBoost model, which supports both regression for predicting Φ and classification for assigning maldistribution severity. The final output blocks emphasize the practical value of the workflow as a fast surrogate for evaluating the studied distributor configurations, reducing repeated CFD and experimental effort within the investigated design space while supporting preliminary thermal-fluid assessment.

The proposed ML model is therefore positioned as a rapid preliminary decision-support tool for screening distributor configurations, while detailed CFD remains necessary for resolving pressure contours, secondary-flow structures, velocity profiles, and full field-level thermal-fluid behavior. The current study is limited by the relatively small size of the available experimental-simulation dataset, the use of published secondary data rather than newly generated laboratory measurements, and the restricted number of distributor geometries, flow-rate levels, and CFD-supported cases. In addition, the CFD-related values used for model comparison were available only for selected configurations and approaches, which limits the generality of the trained AI model across wider hydraulic and thermal-fluid conditions. Future work should therefore expand the dataset using additional 3D-printed distributor designs, larger flow-rate ranges, direct CFD outputs, pressure-drop data, temperature-field measurements, and heat-transfer indicators such as Nusselt number or thermal resistance. Further development may also include physics-informed learning, uncertainty quantification, and digital-twin implementation for rapid screening of water-flow and compact thermal-fluid systems.

5 Conclusions

This study developed an AI-assisted framework for predicting flow maldistribution in 3D-printed water-flow distributors using combined experimental and CFD-supported data. The experimental part of the dataset was obtained from published water-flow measurements of rectangular distributors, conical distributors with obstacle, and conical distributors without obstacle, where the non-uniformity flow coefficient, Φ, was measured under different flow rates. The simulation-supported part was taken from CFD-based comparisons involving selected distributor configurations and VOF/SF modeling approaches, allowing the dataset to include both physical measurements and numerical flow-behavior information. These data were then organized into an AI-ready format and used to develop XGBoost-based regression and classification models. The regression model predicted Φ with strong overall accuracy, achieving R² = 0.962, MAE = 0.126, RMSE = 0.161, MAPE = 4.87%, and CVRMSE = 6.12%. At the configuration level, the lowest errors were obtained for the conical configurations, particularly CD26, with R² = 0.970, MAE = 0.087, and RMSE = 0.113, while higher errors were observed for rectangular distributors due to their larger maldistribution variation. The classification model also performed well in assigning severity classes, achieving accuracy = 92.22%, precision = 92.90%, recall = 92.22%, and F1-score = 92.36%. The confusion matrix confirmed that most errors occurred between neighboring severity classes, with no extreme misclassification between Low and Severe classes. Overall, the results show that XGBoost can provide a fast predictive layer for estimating both the magnitude and severity of flow maldistribution, supporting early-stage screening of distributor designs and reducing reliance on repeated CFD simulations and experimental trials in compact water-flow and thermal-fluid systems.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. During the preparation of this manuscript, the author(s) used ChatGPT, GPT-5.5 Thinking, OpenAI, for language refinement and assistance in improving the clarity of grammar in scientific writing only. The generative AI tool was not used as an author and did not replace the author(s)' scientific judgment, data analysis, interpretation, or final decision-making. All AI-assisted content was reviewed, edited, verified for factual accuracy, and checked by the author(s), who take full responsibility for the final content of the manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Modesto Pérez-Sánchez, Universitat Politècnica de València, Spain

Reviewed by: Abhinav Rajan, University of Pretoria, South Africa

Amjed Abdulelah, University of Kufa, Iraq