An Explainable Physics-Guided Deep Learning Framework for Sustainable Heritage-Building Retrofit and Lifecycle-Carbon Optimisation

Retrofitting heritage buildings promotes sustainable development by balancing decarbonisation with lifetime implications on cultural conservation, comfort, and resource efficiency. Existing data-driven techniques emphasise accuracy over conveying physical-consistency uncertainty, heritage limitations, and embedded operational-carbon trade-offs. This study introduces Adaptive Retrofit and Conservation for Heritage Energy Resilience (ARCHER), an explainable physics-guided deep learning framework that combines reduced-order physics lifecycle-carbon accounting, temporal representation, ensemble uncertainty, and constrained multi-objective optimisation. ARCHER translates dual-target forecasts into retrofit alternatives that are evaluated for moisture legacy impact cost and discomfort before recalculating Pareto solutions using the governing equations. In 8192 Sobol-sampled scenarios for one heritage building, ARCHER attained domain R2 values of 0.9986 for supplied energy and lifespan carbon. Although the Temporal Transformer had lower errors and ARCHER’s intervals remained undercalibrated, the uncertainty ranking lowered RMSE in the least-uncertain decile to 1.51 kWh m−2 year−1 and 15.40 kgCO2e m−2. Exact recalculation kept 42 Pareto optimum possibilities, whereas balanced solutions cut energy and lifetime carbon by 84.3% and 88.1%, respectively, compared to the no-action baseline. ARCHER promotes sustainable heritage decision support by integrating environmental performance, cultural compatibility, and risk screening.