Fresh and mechanical properties of hemp fiber-reinforced concrete: an experimental dosage study with multi-output artificial neural network modelling of mechanical strength

The construction sector’s dependence on carbon-intensive binders and synthetic reinforcement has intensified interest in plant-based fibers as sustainable crack-control reinforcement for concrete. This study experimentally characterises the fresh and 28-day mechanical properties of hemp (Cannabis sativa bast) fiber-reinforced concrete (HFRC) and develops a multi-output Artificial Neural Network (ANN) that predicts its compressive, splitting tensile and flexural strength from fiber dosage. The physical programme comprised eleven dosage levels (0–2.0% by volume in 0.2% steps) of 5% NaOH-treated fiber (aspect ratio ≈ 235) in a fixed M30 base mix (cement 340 kg/m³, water–cement ratio 0.45, superplasticizer 7.48 kg/m³). Slump, compaction factor and Vee-Bee time were measured in triplicate, and three cubes, cylinders and prisms per mix were tested to IS 516; dosage effects were assessed by one-way ANOVA with Tukey and Dunnett comparisons. The 33 physical specimen results per property were kept distinct from 1,100 augmented rows generated from truncated-normal distributions bounded by each mix’s measured mean, standard deviation and range. A 64–32 ReLU multi-output perceptron was evaluated by leave-one-dosage-out cross-validation over ten seeds and benchmarked against polynomial, spline, ridge, tree-ensemble and single-output models. Fiber dosage significantly affected all three strengths ( p < 0.001). Compressive strength peaked at 0.6% (32.8 MPa, + 5.8%), statistically indistinguishable from 0.4 to 0.8%, while splitting tensile and flexural strength peaked at 1.2% (3.95 MPa, + 33.9%; 5.50 MPa, + 31.0%) within statistically equivalent plateaus. On unseen dosages the ANN reproduced the experimental means with RMSE of 0.36, 0.09 and 0.13 MPa, respectively, and outperformed three single-output networks, although a cubic smoothing spline was equally or more accurate. Slump fell from 95 mm to 40 mm and Vee-Bee time rose from 3 s to 20 s across the dosage range, confirming a workability–reinforcement trade-off. The results support hemp fiber dosages of about 0.6–1.2% and show that, for a single-variable dosage study, the multi-output ANN is best used as an exploratory response-curve model rather than as a substitute for physical testing.