In modern commercial building spaces, the smoke generated during indoor fires poses a substantial safety threat to individuals attempting to evacuate. This study leveraged a distributed optical fiber temperature sensing system as the technological foundation. A series of commercial building space models were constructed by employing fire dynamics simulations, and a comprehensive database of simulated fire field information was established. The study built upon this database and proposed a method to establish influence pathways from existing temperature field detection data to three fire risk assessment parameters: carbon dioxide, oxygen, and visibility. The study utilized various machine learning algorithms and focused on selecting and constructing cross-physical field prediction models that optimized the performance of fire risk assessment parameters at the fire key plane. The findings demonstrated that the Multilayer Perceptron (MLP) algorithm model yielded the most accurate fit for carbon dioxide volume fraction, with a Mean Absolute Percentage Error (MAPE) of 3.80% between the model predictions and actual values. The Artificial Neural Network (ANN) algorithm showed superior fitting effects for oxygen volume fraction and visibility, with MAPEs of 1.37% and 3.10%, respectively. Furthermore, the study validated the generalizability of these optimal models by altering spatial scales, revealing robust performance. This study utilized a single optical fiber sensor to achieve dynamic prediction across different physical fields during fire incidents, expanding the application scope of optical fiber sensors in monitoring gas and visibility fields in commercial building fires. This not only enhanced the level of fire monitoring but also effectively reduce equipment costs. It provided effective data support for planning evacuation routes for individuals and rescue arrangements for firefighters, further ensuring personnel safety.