This study presents a refined deep learning–based time-series surrogate model for the BRI2-CRIEPI zone fire model using a Long Short-Term Memory (LSTM) architecture. The surrogate model was trained on a large dataset of simulations generated by the BRI2-CRIEPI model and was designed to predict the temporal evolution of key fire behavior parameters from static initial conditions. Twenty influential input parameters, representing compartment geometry, fire source characteristics, ventilation conditions, wall thermal properties (including wall density), and initial conditions, were identified and used for model training.

The proposed surrogate model consists of a three-layer LSTM network with 256 hidden units per layer and predicts six output variables characterizing fire behavior, including heat release rate (HRR), smoke layer temperature, smoke layer height (upper layer thickness), compartment pressure, and oxygen concentration. Notably, the model can forecast the complete time histories of these outputs without requiring time-stepped input data. To enhance training stability and convergence, a teacher forcing strategy was employed.

Using a dataset of 100,000 simulated fire scenarios, the LSTM surrogate model achieved high predictive accuracy on unseen cases, with an average normalized mean absolute error (nMAE) of approximately 0.029 across all output parameters, improved from 0.042 in the previous configuration. In addition, the surrogate model was compared with full-scale experimental HRR data from mechanically ventilated single-compartment tests (PRS-SI-D1 and PRS-SI-D2) in the OECD/NEA PRISME project. Although the early growth and burning duration showed discrepancies due to unlearned variations in fire growth rate and fuel amount, the surrogate reproduced the quasi-steady HRR levels reasonably well.