Physics guided Transformer XL framework for real time cross building fire spread simulation

In urban high-density building clusters, fires not only tend to spread rapidly within individual buildings but can also breach building boundaries and propagate to adjacent structures. Traditional field simulation methods suffer from high computational costs and significant response delays, while existing data-driven models struggle to capture long-range dependencies across multiple buildings. Transformer-XL, with its segment-level recursive mechanism and relative position encoding, offers unique advantages for long-sequence modeling tasks. Therefore, this study proposes a Transformer-XL-based framework for real-time cross-building fire spread simulation. The framework constructs a multimodal input layer to uniformly encode numerical monitoring data, spatial topology data, and meteorological data, and combines local perception and global dependency capture capabilities through an improved Transformer-XL encoder. Experiments are conducted on a large-scale synthetic dataset generated by FDS (Fire Dynamics Simulator), containing 120 independent fire scenarios with 8–15 buildings each. The dataset is split into training, validation, and test sets in a 7:1:2 ratio, with test scenarios being completely independent of training scenarios. Model performance is evaluated using mean Intersection over Union (mIoU), Mean Absolute Error (MAE), inference accuracy, and average response time. Experimental results show that the proposed model achieves a maximum mIoU of 51%, maintaining an accuracy above 48% throughout the 16-minute prediction horizon. The model achieves an inference accuracy exceeding 94% across all tested building spacings, with a peak accuracy of 97.8% at 20 m spacing, improving accuracy by 8.8, 5.2, and 2.7% points compared to LSTM, CNN-LSTM, and the basic Transformer, respectively. The average response time is 160ms, and the data fit rate reaches 98.7%. When building density increases from 0 to 3, the model’s MAE (9.5→31.7) is the lowest among all models. These results establish a computationally efficient benchmark under controlled FDS-generated conditions and support further validation of physics-guided long-sequence models for cross-building fire spread forecasting.

Authors

Institutions

Publication Details

Journal
Discover Internet of Things
Published
2026-11-17
DOI
https://doi.org/10.1007/s43926-026-00555-9
Primary Topic
Fire dynamics and safety research
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Physics guided Transformer XL framework for real time cross building fire spread simulation

Jing Song, Yajun Nie, Yunfang Lu, Qiong Zhang et al.
5 links
Discover Internet of Things
Fire dynamics and safety research
article

Physics guided Transformer XL framework for real time cross building fire spread simulation

Jing Song, Yajun Nie, Yunfang Lu, Qiong Zhang, Yanqiong Zhang, Jinfeng Qin
article en

Abstract

In urban high-density building clusters, fires not only tend to spread rapidly within individual buildings but can also breach building boundaries and propagate to adjacent structures. Traditional field simulation methods suffer from high computational costs and significant response delays, while existing data-driven models struggle to capture long-range dependencies across multiple buildings. Transformer-XL, with its segment-level recursive mechanism and relative position encoding, offers unique advantages for long-sequence modeling tasks. Therefore, this study proposes a Transformer-XL-based framework for real-time cross-building fire spread simulation. The framework constructs a multimodal input layer to uniformly encode numerical monitoring data, spatial topology data, and meteorological data, and combines local perception and global dependency capture capabilities through an improved Transformer-XL encoder. Experiments are conducted on a large-scale synthetic dataset generated by FDS (Fire Dynamics Simulator), containing 120 independent fire scenarios with 8–15 buildings each. The dataset is split into training, validation, and test sets in a 7:1:2 ratio, with test scenarios being completely independent of training scenarios. Model performance is evaluated using mean Intersection over Union (mIoU), Mean Absolute Error (MAE), inference accuracy, and average response time. Experimental results show that the proposed model achieves a maximum mIoU of 51%, maintaining an accuracy above 48% throughout the 16-minute prediction horizon. The model achieves an inference accuracy exceeding 94% across all tested building spacings, with a peak accuracy of 97.8% at 20 m spacing, improving accuracy by 8.8, 5.2, and 2.7% points compared to LSTM, CNN-LSTM, and the basic Transformer, respectively. The average response time is 160ms, and the data fit rate reaches 98.7%. When building density increases from 0 to 3, the model’s MAE (9.5→31.7) is the lowest among all models. These results establish a computationally efficient benchmark under controlled FDS-generated conditions and support further validation of physics-guided long-sequence models for cross-building fire spread forecasting.

Discover Internet of ThingsVol. 6(1)
Anhui Institute of Architectural Research and Design (CN), Hefei University of Economics (CN)
Sustainable cities and communities, Industry, innovation and infrastructure
Fire dynamics and safety research
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon