Optimized LSTM Model Using Simulated Annealing for Autoignition Temperature (AIT) Prediction as a Hazard Indicator

Abstract

Autoignition temperature (AIT) is the minimum temperature at which a substance sparks spontaneously in air under normal atmospheric pressure without an external ignition source, such as a flame. This parameter is crucial for industrial safety, particularly in the production, processing, handling, transportation, and storage of flammable materials. However, conventional AIT measurement methods are time-consuming, expensive, and carry significant risk. As an alternative, in silico approaches based on machine learning can be used to develop AIT prediction models. Among these approaches, Long Short-Term Memory (LSTM) networks are particularly effective for modeling complex non-linear relationships. However, the performance of LSTM models is highly sensitive to the configuration of numerous hyperparameters, making manual tuning inefficient. Consequently, an automated optimization strategy is required to identify the optimal model architecture. This study aims to develop an AIT prediction model as a hazard indicator using the Long Short-Term Memory (LSTM) method optimized with Simulated Annealing (SA). Experimental results demonstrated that the proposed SA-LSTM Model with a cooling schedule of ΔT = 0.7 outperformed the unoptimized baseline architecture. The optimization process improved the R² on the data test from 0.5682 to 0.5939 and reduced the RMSE from 74.35 K to 72.10 K. Furthermore, the MAPE decreased from 9.29% to 8.87%. These findings confirm that the SA optimized LSTM model provides a more reliable and robust hazard indicator.