Developing a deep learning–based model to predict simulated building performance using the spatial layout of residential apartments in Isfahan

Spatial configuration is a primary determinant of residential building performance, yet evaluating trade-offs across design alternatives remains computationally intensive. This study introduces an integrated computational framework combining a Generative Spatial Layout (GSL) algorithm, Building Performance Simulation, and a hybrid multi-input Deep Learning (DL) model to rapidly predict multi-objective performance. Focusing on south-facing rectangular apartments in Isfahan, 27,000 distinct design configurations were generated and evaluated across six performance metrics: cooling demand, heating demand, useful daylight illuminance (UDI), embodied carbon, operational carbon, and embodied water. The dataset was partitioned into training (70%), validation (10%), and testing (20%) subsets to train the hybrid DL architecture, which integrates dense numerical layers and Transformer-based spatial sequence processing. The Mean Squared Error (MSE) value demonstrates the optimal performance of the study’s ML model, indicating high prediction accuracy. Moreover, the coefficients of determination (R 2 ) for all performance indicators range between 0.95 and 0.997, demonstrating strong predictive performance within the evaluated dataset. Analysis further reveals the significant influence of the dimensions, location, and number of Indoor Spaces, particularly the living room and Vertical Circulation Core, on energy and environmental performance. The hybrid DL approach proves effective for predicting residential building performance using multidimensional GSL datasets. Findings identify spatial optimization and strategic Vertical Circulation Core placement as critical factors for improving overall building efficiency.