An adaptive deep neural network-based assessment framework for environmental-cost-mechanical performance of engineered cementitious composites

The adoption of new construction materials requires quick evaluation of their environmental-cost-mechanical performance adequacy. However, comprehensive life cycle assessment (LCA) and cost analysis are often complex and time-consuming, while experimental characterization of mechanical properties typically requires days or even weeks. This study employs an engineering-oriented data-driven surrogate framework through deep neural network (DNN) to predict environmental impacts, cost, and mechanical properties of an advanced bendable concrete, engineered cementitious composites (ECC). The mechanical properties, including compressive strength, tensile strength, and tensile strain capacity, were collected from reference for 358 groups of ECC. Then, environmental impacts and cost databases were developed through our “cradle-to-gate” LCA and material cost analysis, with functional unit as 1 m 3 of ECC material. A DNN is first trained, and an adaptive DNN (ADNN) with dynamic loss weighting is introduced to balance multiple targets and improve generalization. The results show that compared to the DNN model, ADNN model reduced test-set MSE by 44.99% and MAE by 4.69%, reflecting improved convergence, generalization, and robustness, under realistic data constraints. Feature attribution analysis identifies Portland cement, polymer fiber content, and fiber class as dominant predictors. The proposed framework provides a fast, automated alternative to traditional LCA, mechanical testing, and cost analysis approaches, enabling a comprehensive evaluation of mechanical properties, cost, and environmental impacts, and supporting intelligent design and promotion of advanced ECC materials.