Bayesian neural network framework for uncertainty-aware design of fiber-reinforced cementitious composites

Performance prediction and design optimization of fiber-reinforced cementitious composites (FRCCs) are challenged by sparse and noisy experimental data, limiting predictive accuracy and design robustness due to inadequate uncertainty quantification. The lack of systematic uncertainty quantification methods compromises predictive accuracy and design robustness. This study develops a Bayesian framework integrating Bayesian neural networks (BNNs) with multi-objective optimization. The framework establishes probabilistic composition-property mappings for mechanical properties and flowability, explicitly quantifying aleatoric and epistemic uncertainties. Based on such BNNs framework, an integrated hybrid TOPSIS decision model, which combines NSGA-II multi-objective optimization with quantified uncertainty, is established. The framework’s efficacy is validated through experimental testing. This research provides civil engineers with a practical framework for designing FRCCs under data scarcity conditions. The proposed BNNs approach quantifies prediction uncertainties, enabling risk-aware material design. The integrated optimization methodology balances conflicting performance objectives while considering reliability, supporting the development of robust FRCCs formulations for structural repair and strengthening applications. The framework’s modular nature allows adaptation to other civil engineering materials with limited experimental data.