Screen hundreds of formulations in days instead of months
At the intersection of chemistry and machine intelligence, researchers at IMDEA Materials Institute have trained an AI to foresee how well epoxy resins will resist fire — a question that has long demanded costly, repetitive laboratory work to answer. By learning from 510 material samples, the model can now read a proposed formulation and predict its fire behavior before a single experiment is run. In a world where polymers quietly underpin aircraft, electric vehicles, and buildings, the ability to anticipate danger at the design stage rather than discover it in flames carries a quiet but profound significance.
- Epoxy resins are indispensable to modern industry yet carry an inherent vulnerability — they burn, and that flammability has long constrained where and how safely they can be used.
- Traditional flame-resistance testing demands that materials be synthesized and ignited repeatedly, a process that drains resources and slows the path from promising compound to certified product.
- A machine-learning model trained on 510 phosphorus-retardant epoxy samples now predicts two standardized fire-safety metrics — UL-94 rating and Limiting Oxygen Index — directly from molecular structure.
- A four-tier classification system translates the AI's predictions into plain guidance — excellent, good, moderate, or poor — giving engineers actionable direction without requiring deep expertise in fire science.
- Validated against external case studies, the framework is already pointing toward faster development of safer materials for electric vehicle batteries, aircraft interiors, electronics, and construction.
At IMDEA Materials Institute, researchers have built an AI system capable of predicting how well epoxy resins resist fire — a development that could fundamentally change how safer industrial materials are designed. The work, published in Polymer Degradation and Stability, bypasses the traditional cycle of synthesizing a compound, testing it under controlled ignition, and iterating — a process that is slow, expensive, and sensitive to experimental variability.
Epoxy resins are foundational to construction, automotive engineering, electronics, and aerospace, yet they burn. The industry's answer has been phosphorus-based flame-retardant additives, which offer effectiveness without the environmental and health concerns associated with halogen compounds. Dr. Qiong Tan, a postdoctoral researcher at IMDEA, led the training of a machine-learning model on 510 epoxy composite samples incorporating these additives in varied configurations. The model analyzes molecular structure, mixing ratios, and other formulation variables to predict two critical fire-safety metrics: the UL-94 vertical flammability rating and the Limiting Oxygen Index.
The UL-94 test measures how a material behaves after an ignition source is removed — whether it keeps burning, glows, or drips. The LOI identifies the minimum oxygen concentration needed to sustain combustion. Together, they offer a far richer picture of fire behavior than either metric alone. Crucially, the AI's output feeds into a unified classification framework that sorts materials into four tiers — excellent, good, moderate, or poor — giving engineers clear, actionable guidance on whether a formulation merits further development.
Already validated against external case studies, the system is proving robust in real-world scenarios. IMDEA's High-Performance Polymers and Fire Retardants Research Group plans to expand the database to cover additional polymer types and flame retardants. The practical stakes are immediate: safer electronics, more thermally stable EV batteries, fire-code-compliant aircraft interiors, and more protective building materials — all reachable faster, and with greater confidence, than before.
At IMDEA Materials Institute, researchers have built an artificial intelligence system that can predict how well epoxy resins will resist fire—a capability that could reshape how industries develop safer materials. The work, published in Polymer Degradation and Stability, sidesteps the traditional path of designing a compound, synthesizing it in the lab, and then testing it repeatedly under controlled conditions, a process that consumes time and money and remains vulnerable to the quirks of experimental setup.
Epoxy resins are everywhere in modern manufacturing. Construction sites use them. Automotive engineers rely on them. Electronics makers and aerospace companies depend on them. Yet these polymers have a fundamental weakness: they burn. This flammability limits where they can be deployed, especially in applications where fire safety is non-negotiable. The industry's response has been to add flame-retardant additives to the resin formulations. Among these additives, phosphorus-based flame retardants have emerged as the most promising option, partly because they work without halogen compounds, which carry their own environmental and health concerns.
Dr. Qiong Tan, a postdoctoral researcher at IMDEA, led the effort to train a machine-learning model on data from 510 epoxy composite samples, each incorporating phosphorus-based flame retardants in different configurations. The model learned to analyze the molecular structure of the flame retardants themselves, the way they were mixed into the epoxy formulation, and other variables that influence how the material behaves in a fire. From this analysis, it can now predict two critical measures of fire resistance: the UL-94 vertical flammability rating and the Limiting Oxygen Index, or LOI.
The UL-94 test is a standardized protocol that exposes polymeric materials to a controlled ignition source and measures how long the material continues to burn after the flame is removed, whether it glows, and whether it drips molten material. The LOI, by contrast, determines the minimum concentration of oxygen needed in an oxygen-nitrogen mixture to sustain combustion of the material. Neither metric alone tells the complete story of how a material will perform in a real fire. Together, they provide a much fuller picture. The innovation here is not just the prediction itself but the framework that wraps around it. The AI results feed into a unified classification system that sorts materials into four semantic categories: excellent, good, moderate, or poor. Engineers and materials designers can now look at a proposed formulation and receive clear, actionable guidance on whether it is worth pursuing further.
The model has already been tested against external case studies and has proven robust enough to handle real-world scenarios. IMDEA Materials, through its High-Performance Polymers and Fire Retardants Research Group led by Prof. De Yi Wang, is now planning to expand the database to encompass other types of polymers and flame retardants beyond epoxy. The immediate applications are concrete and consequential: electronic components could become safer, electric vehicle batteries more resistant to thermal runaway, aircraft interiors more compliant with fire codes, and building materials more protective in the event of a fire. What was once a slow, expensive, trial-heavy process can now be accelerated, allowing safer materials to reach the market faster and with greater confidence.
Notable Quotes
The traditional development of efficient flame retardants involves a design, synthesis and laboratory testing process that is slow, costly and highly sensitive to experimental conditions.— Dr. Qiong Tan, IMDEA Materials Institute
Direct applications could improve the safety of electronic components, electric vehicle batteries, aircraft interiors and construction materials.— Dr. Qiong Tan
The Hearth Conversation Another angle on the story
Why does it matter that this model can predict fire resistance instead of just testing materials the old way?
Because the old way—design, synthesize, test, repeat—takes months and costs a lot. You're paying for lab time, equipment, skilled technicians. With this model, you can screen hundreds of formulations computationally in days. You only build and test the most promising candidates.
But how does the model actually know what will burn and what won't? It's trained on 510 samples. That seems small.
It's trained on the right 510 samples. Each one has detailed molecular data about the flame retardant, how it was mixed, and two standardized fire-safety measurements. The model learns the patterns in that data—which molecular structures tend to perform well, which formulations work synergistically. It's not guessing. It's pattern recognition at the molecular level.
The framework sorts materials into four categories. Doesn't that oversimplify things? Real fire behavior is complicated.
It does simplify, but that's the point. Engineers don't need infinite precision; they need clarity. Is this material excellent or poor? That distinction is actionable. And the model is still using two different fire-safety metrics underneath, so it's not throwing away nuance—it's organizing it into language people can actually use.
What happens next? Is this ready to deploy in industry?
It's been validated on external cases, so it's already being used. But the real expansion is coming—they're building out the database to cover other polymers and flame retardants beyond epoxy. The more data the model sees, the more reliable it becomes. In five years, this could be standard practice in materials development.