UNIVERSITY PARK, Pa. — The mining industry is shifting toward battery-powered equipment to improve efficiency, but there is risk associated with lithium-ion batteries due to their high energy density. They can fail, overheat and, in some cases, catch on fire, releasing toxic emissions. By combining three modeling approaches to build virtual mining systems, Penn State researchers now more comprehensively understand how to predict and mitigate lithium-ion battery fires in mines before they spread, endangering miners and halting operations.
The team used computational fluid dynamics (CFD) modeling, which enables better understanding of how dynamic materials move around structures, to build virtual underground room-and-pillar mine environments — one of the most common types of mine systems where deposits, such as coal, are excavated in a grid that leaves large, unmined columns of the material to support the weight of the overhead rock. They then trained deep learning models to better understand how temperature, smoke and carbon monoxide evolve under different ventilation conditions in the virtual mines. The findings, published in Process Safety and Environmental Protection, have immediate implications on how to deploy automated control systems to promote safety and operational resilience in mining systems electrified by lithium-ion batteries, according to the researchers.
Khadija Omar Said, a doctoral student in the John and Willie Leone Family Department of Mineral and Energy Engineering and lead author on the paper, said that integrating battery-electric vehicles (BEVs) with physics-informed artificial intelligence and automated control systems is essential for the next frontier of mining.
“Mining companies are looking to optimize operations as much as possible,” Said said. “They see the potential to increase output and safety through automation supported by battery-powered systems.”
Population growth, clean energy transitions and the spread of technology are driving demand for minerals, pushing the mining industry to rethink how operations are designed and managed, according to the research team, which also included Ashish Ranjan Kumar, assistant professor of mining engineering, and Sekhar Bhattacharyya, associate teaching professor of mining engineering.
“Mounting evidence shows the benefits of battery-powered machines, with recent studies showing BEVs can transport up to significantly more material than traditional internal combustion engine vehicles,” Kumar said. “Studies also show significant reductions in heat, noise levels and generation of noxious fumes like diesel exhaust and particulate matter when BEVs are used, which lowers operating costs because ventilation needs decrease.”
Lithium batteries are commonly used for their high energy efficiency, Said explained. But they require strict operation because they carry an inherent danger of thermal runaway — an uncontrollable, explosive chain reaction that occurs when the internal temperature of the battery rises faster than its internal threshold.
Because of the battery's chemical nature, the self-accelerating reaction simultaneously produces both fire and oxygen, which in turn feeds the fire. In the confined, limited-ventilation spaces of underground mines, fumes and heat pose an even greater danger, according to the researchers.
“Once thermal runaway starts, there's no stopping it — you can only try to reduce the consequences,” Said said.
Numerous causes can compromise a lithium battery and lead to thermal runaway, including physical impact to the battery, environmental factors such as ambient temperature, mismanagement leading to repeated overcharging or simple manufacturing defects.
“Making matters worse, lithium battery cells are commonly tightly packed together to produce more power, and a thermal reaction in one cell can quickly trigger a chain reaction across the other cells,” Said said. “The key is prevention and developing better ways to predict it. We need to understand the signatures of thermal runaway, Then, if we know the signs, maybe we can either prevent the initial ignition or mitigate that cascading chain reaction.”
Advances in AI and simulations have provided a safe way to test thermal runaway scenarios and emulate adjustments to ventilation systems. The research team used the Fire Dynamics Simulator, a CFD modeling tool, to simulate thermal runaway in an underground mine and generate temperature, smoke and carbon monoxide data at multiple monitoring locations over time. They then tested how different ventilation measures, such as fan speed, would affect the circulation and spread of toxic fumes. The simulations let researchers see how hazardous levels could drop or, alternatively, how small tweaks could create pockets of toxic air downstream of the initial incident.
“Despite the advances in modeling and the increase in the number of sensors deployed throughout a modern mining operation, limited studies have developed autonomous emergency response frameworks for battery fires,” Said said. “We’re bridging this gap and hopefully laying the foundation for more economical solutions.”
According to the researchers, a more economical real-time approach matters because no two mines are the same. Differences in mine geometry, ventilation networks and equipment can influence how heat, smoke and gases move throughout a mine. Intelligent mining systems must account for these changes while providing predictions quickly enough to support real-time monitoring and control.
High-fidelity CFD simulations can be computationally intensive for real-time applications, so the researchers developed deep learning surrogate models that learn the complex thermal dynamics captured by the simulations while providing much faster predictions. This approach could support real-time monitoring and, ultimately, automated control in intelligent mining systems.
The researchers evaluated three deep learning models that were each designed to process data differently, Said explained, with each having distinct advantages and shortcomings for prediction. A common test was to feed the models a data set captured from a real mine or thermal runaway incident and observe how the models would play out the scenario in the virtual mine.
“Basically, we test the models by asking them to solve problems we already have the answer to,” Said said. “If the models consistently reproduce values they were not shown during training, that gives us confidence in their ability to predict new scenarios within a reasonable range.”
The results showed that increasing airflow enhanced heat removal and diluted smoke and toxic gases, improving conditions enough for miners to escape safely. However, increasing airflow too much produced limited additional benefits as fire-induced movement increased and competed with the ventilation-driven flow.
Each deep learning model showed distinct strengths, depending on conditions such as how the airflow changes were administered or spatial location. Together, they formed a comprehensive understanding of how different deep learning architectures can capture the evolving conditions in a lithium battery thermal runaway event.
Based on the simulations, the researchers also identified data opportunities for improved sensor placement and data reconstruction. Said explained that this work revealed an opening to design a blueprint for intelligent monitoring and control — determining where sensors are needed, reconstructing missing information, and identifying which mitigation measures to deploy and when.
“We've demonstrated that we can predict the impact of thermal runaway in a confined mine environment under unique ventilation conditions; the next step is to actually control it,” Said said. “Now that we understand the dynamic conditions that cause thermal runaway, we should be able to use this data to develop control algorithms that can automatically trigger the fans or sprinklers in a way that can suppress the effects and ensure resilience in mining operations.”
This research was partially supported by the National Institute for Occupational Safety and Health under grant number 75D30124C20331 and by the Thomas V. and Jean C. Falkie Fellowship in the College of Earth and Mineral Sciences. Support was also provided by Penn State’s Institute for Computational and Data Sciences. This content is the sole responsibility of the authors and does not necessarily reflect the views of the funders.