Diffusion-based generative AI framework for long-term degradation forecasting and risk assessment of steel fibre reinforced fly ash-based concrete
Vikrant S. Vairagade
Abstract
Reliable lifetime estimations are a prerequisite for durable transport and energy networks such as steel fibre reinforced fly ash (SFRC-FA) concrete, when subjected to combined ambient conditions such as chloride, carbonation, sulfate and thermal cycles. Conventional empirical equations and recurrent networks treat each stressor in isolation while ignoring space geometry and holding on to sparse long-term records, giving high safety margins and expensive over-designs. To close these gaps, the present study introduced a diffusion-based generative AI framework that unified long-horizon forecasting, synthetic data generation and multi-criteria risk aggregations. The Time-Entropy Residual Fusion Transformer (TERFT) couples’ entropy-weighted residual blocks with temporal transformers, cutting 60-month forecasting RMSE by 18% compared with long short-term memory / gated recurrent unit (LSTM/GRU) while preserving 85% confidence bounds. Over 10 000 stressor-conditioned degradation trajectories were synthesised using physics informed multi-diffusion generative modelling simulator (MDGS), with < 7% deviation from 10–15 years field records, enriching scarce datasets. A Graph Integrated Structural Fusion Encoder (GISFE) incorporates topology of beams and slabs, sensor layout, and exposure mapping, improving prediction accuracy by 20% location-wise. Outputs of forecast then find applications in the Multi-Criteria Fracture Risk Aggregator (MCFRA), lumping volume, velocity of degradation with severity of stressors and uncertainty into adaptive risk tensor format with 92% accuracy in terms of failure classification and lead delays of 18–24 months. Latent Feature Attribution via Diffusion Influence Scoring (LFADIS) reverse-traces all predictions ideally latent to environmental and design vectors, lowering loss in attribution by 21% while enhancing transparency. The pipeline would then give accurate lifetime predictions, create rare tail-scenario data, and provide understandable risk scores to optimise cost in design and proactive maintenance scheduling for sustainable concrete infrastructure across climatic zones and service conditions within global portfolios.