AI Predicts Uneven Water Flow in 3D-Printed Distributors Before Testing

When a single stream of water must be split evenly among many parallel channels, the humble distributor sitting at the heart of the system quietly decides whether everything runs smoothly or descends into hydraulic chaos. Uneven flow sharing, known as flow maldistribution, can starve some channels of water while flooding others, upsetting pressure balance and destabilizing the convective heat transfer that compact cooling plates, heat exchangers, and parallel-channel thermal devices depend on. For engineers, that means hotspots, uneven cooling, and reduced reliability in everything from microchannel heat sinks to industrial water systems. Now a team of researchers in Iraq has built an artificial intelligence framework that predicts how badly a given distributor will maldistribute water before anyone runs expensive simulations or laboratory trials, and the results suggest that machine learning could become a standard first filter in thermal-fluid design.

The study, published in Frontiers in Water by researchers at the University of Technology-Iraq and Al-Mustansiriyah University in Baghdad, tackles a familiar bottleneck in early-stage engineering. Computational Fluid Dynamics, the workhorse simulation tool for studying how fluids move through hardware, delivers rich detail but demands considerable setup time, meshing effort, and computational cost, especially when dozens of geometries and operating conditions must be compared. Physical experiments are even slower. The research team, led by Luttfi A. Al-Haddad, therefore turned to Extreme Gradient Boosting, or XGBoost, an ensemble machine learning algorithm that combines many weak decision-tree learners sequentially to capture non-linear relationships while regularization keeps overfitting under control. Before settling on XGBoost, the authors compared it against alternative models and found it delivered the best overall prediction accuracy, error reduction, and classification stability for their dataset.

What makes the new framework distinctive is its data foundation. Rather than training purely on simulations, the model was built from a combined experimental and CFD-supported dataset drawn from published measurements of 3D-printed water-flow distributors. The experimental portion came from water-flow tests of three distributor families: rectangular manifold-type distributors and conical distributors, the latter both with and without an internal obstacle, each manufactured at heights of 13, 26, and 52 millimeters and tested across flow rates from 100 to 400 milliliters per minute. The simulation portion added Volume of Fluid (VOF) and single-phase flow results for selected conical configurations. Five input variables—distributor family, height, flow rate, data source, and simulation method—fed the model, while the prediction target was the non-uniformity flow coefficient, Φ, a percentage that quantifies just how far the flow sharing deviates from perfect uniformity.

The experimental baseline itself tells a striking geometric story. The worst performer, a shallow rectangular distributor labeled RD13, produced a maldistribution coefficient of 34.22 percent at the lowest flow rate of 100 milliliters per minute—meaning flow sharing was wildly uneven—before improving to 2.04 percent at 400 milliliters per minute. The taller RD52 began at 13.96 percent and also improved with flow rate. The conical geometries behaved far better: the CDO26 distributor with an internal obstacle varied only between 0.94 and 2.13 percent across all tested flow rates, while the smooth CD13 cone stayed between 0.63 and 2.30 percent. Averaged over all conditions, the best distributors, including CDO13, CDO52, and CD26, achieved average non-uniformity coefficients below one percent, whereas RD13 averaged a troubling 8.236 percent with a large standard deviation of 5.636. The lesson is clear: tapered, conical flow paths share water far more stably than rectangular manifolds, particularly at low flow rates where inertial forces are weak.

The CFD comparison added an important layer of nuance. Simulations using the Volume of Fluid approach captured the strong low-flow maldistribution of the tallest conical configuration, CDO52, much better than single-phase modeling, which produced smoother and generally lower predictions of Φ. For CDO52, the average experimental coefficient was 1.98 percent and the VOF average was a close 1.92 percent, while single-phase flow predicted only 0.73 percent. The authors attribute discrepancies between measurement and simulation to real-world effects that experiments capture but models idealize: 3D-printing tolerances, surface roughness, flow disturbances, and measurement variability on one side, and model assumptions, boundary conditions, and numerical formulation on the other. Rather than treating those differences as a problem, the team deliberately trained the AI on both data sources so the model would learn the realistic range of possible flow behavior instead of a single idealized version of it.

The machine learning architecture pairs regression with classification in a two-stage decision framework. The regression model predicts Φ as a continuous output, expressed as the sum of contributions from fifty regression trees, each with a maximum depth of two and a conservative learning rate of 0.05, alongside L1 and L2 regularization to suppress overfitting on the limited 90-sample dataset. A separate multiclass XGBoost classifier, using the multi:softprob objective, then assigns each design to one of four pragmatic severity categories defined by the authors: Low for Φ at or below 1 percent, Moderate between 1 and 5 percent, High between 5 and 10 percent, and Severe above 10 percent. These thresholds were introduced as engineering classification boundaries for comparative evaluation rather than adopted from any international standard, but they convert an abstract coefficient into an immediately actionable design verdict.

The performance numbers are impressive for such a compact dataset. Across the full combined dataset, the regression model achieved a coefficient of determination of 0.962, meaning it explained over 96 percent of the variance in maldistribution, with a mean absolute error of 0.126, a root mean square error of 0.161, a mean absolute percentage error of 4.87 percent, and a coefficient of variation of the RMSE of 6.12 percent. The best configuration-level performance came from the smooth CD26 cone, with an R² of 0.970 and MAE of 0.087, while the erratic RD13 rectangle proved hardest to fit, with an MAE of 0.185. Notably, the conical distributors yielded lower prediction errors across the board, consistent with their more stable flow behavior. The classifier matched this success, achieving 92.22 percent accuracy, 92.90 percent precision, 92.22 percent recall, and an F1-score of 92.36 percent.

Perhaps most encouraging is how the model failed. In the confusion matrix, it correctly classified 38 Low, 36 Moderate, 6 High, and 3 Severe samples, and every error landed in an adjacent severity class—never did a Low design get labeled Severe or vice versa. The predicted class distribution deviated from the actual distribution by only a sample or two in each category. To test generalizability beyond the fitting stage, the team evaluated a 50-sample holdout subset restricted to the Moderate class, chosen deliberately because it sits between the Low and High decision boundaries and therefore poses a sensitive discrimination test. The trained classifier correctly identified 47 of the 50 unseen samples as Moderate, an external accuracy of 94 percent, with the three errors again confined to neighboring classes and no sample mistaken for Severe.

The practical payoff is speed. Instead of running repeated CFD simulations or printing and testing distributor after distributor, engineers can use the trained model as a rapid screening layer that estimates both the magnitude and the severity of maldistribution for candidate geometries, then reserve detailed CFD and physical testing for the most critical or borderline cases. Detailed simulation remains essential for resolving pressure contours, secondary-flow structures, and velocity profiles that a tree-based surrogate cannot see, but the AI layer trims the brute-force search dramatically. The authors frame the work as a decision-support tool that maintains attention to pressure balance, flow uniformity, and stable convective heat transfer while accelerating preliminary thermal-fluid assessment.

The study also candidly acknowledges its limits. The dataset is small, built from published secondary measurements rather than newly generated laboratory data, and covers a restricted set of geometries, flow rates, and CFD cases; simulation values were available for only selected configurations, limiting the model’s generality across wider hydraulic conditions. Future work, the authors suggest, should expand the dataset with additional 3D-printed designs, broader flow-rate ranges, direct CFD outputs, pressure-drop data, temperature-field measurements, and heat-transfer indicators such as the Nusselt number, potentially incorporating physics-informed learning, uncertainty quantification, and digital-twin implementations. Even so, the message resonates well beyond water distributors: with the right combined experimental and simulation data, machine learning can compress weeks of simulation-and-prototype iteration into seconds of inference, offering engineers a fast, physics-aware first look at whether their design will share its flow evenly—or send one channel starving while another drowns.

Subject of Research: AI-based prediction of flow maldistribution in 3D-printed water-flow distributors using combined experimental and CFD data

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

Article References: Prediction of flow maldistribution in 3D-printed water-flow distributors for thermal-fluid system design. (n.d.). https://doi.org/10.3389/frwa.2026.1921076

Image Credits: AI Generated

DOI: 10.3389/frwa.2026.1921076

Keywords: flow maldistribution, 3D printing, water-flow distributors, XGBoost, machine learning, computational fluid dynamics, thermal-fluid systems, heat transfer, non-uniformity coefficient, Volume of Fluid, design screening, hydraulic engineering