The material at the heart of the study is a hybrid nanocomposite that combines epoxy resin with two very different nanoscale fillers: multi-walled carbon nanotubes, or MWCNTs, and nanoclay. Each filler brings a distinct talent. Carbon nanotubes form percolating conductive networks that allow the coating to carry electrical current and respond to changes in temperature or chemical exposure. Nanoclay platelets, when properly dispersed into intercalated or exfoliated structures, act as physical barriers that slow heat transfer and the diffusion of combustible gases, while also stiffening the polymer matrix. The challenge has always been that these benefits do not scale independently. Adding more of one filler can disrupt the dispersion of the other, shift the curing behavior of the epoxy, or push one property past its optimum while another collapses.

Traditionally, researchers have navigated this compositional space through trial and error, or through statistical design-of-experiments methods such as response surface methodology. The Vietnamese team adopted a four-factor Box–Behnken design, an efficient experimental layout that samples the corners and center of a multi-dimensional formulation space without requiring every possible combination to be tested. That design was coupled with response surface modeling, but the authors went further by training a Random Forest model, an ensemble machine learning method that builds many decision trees on randomized subsets of the data and averages their predictions. The Random Forest approach substantially outperformed classical response surface methodology, achieving coefficients of determination, or R² values, of 0.986 for limiting oxygen index, 0.995 for tensile strength, and 0.998 for the logarithm of electrical conductivity.

Those numbers matter because they describe how confidently the model can predict real material behavior from formulation variables alone. An R² approaching 1.0 means nearly all of the variation in the measured property is captured by the model. In practical terms, the researchers could ask the algorithm what happens if the nanotube loading rises while the nanoclay fraction falls, and receive a reliable answer without mixing a single new batch of resin. To make the model interpretable rather than a black box, the team applied SHAP analysis, a technique borrowed from explainable artificial intelligence that quantifies how much each input factor contributes to each prediction, including the direction and nature of its influence across the formulation range.

The optimization itself relied on multi-objective desirability functions, a strategy that seeks a balanced formulation window rather than maximizing any single property in isolation. This distinction is central to the study’s philosophy. A coating with record-breaking flame retardancy but brittle mechanics would fail in service, just as a tough but flammable coating would fail certification. By defining desirability across fire performance, mechanical strength, conductivity, and sensing response simultaneously, the framework identified a composition where all properties land in an acceptable, mutually compatible zone. The authors report that this machine learning-assisted workflow reduced the experimental burden by approximately 50 to 65 percent compared with a full-factorial exploration of the same design space, a substantial saving in laboratory time, materials, and cost.

The optimized coating delivered impressive measured performance. Its limiting oxygen index, the minimum oxygen concentration in an atmosphere that sustains candle-like burning, reached 28.0 plus or minus 0.2 percent, a threshold associated with genuinely flame-retardant behavior. In cone calorimetry testing, the peak heat release rate was reduced to 720 plus or minus 9 kilowatts per square meter, indicating that the material releases heat far more slowly when exposed to fire. Mechanically, the coating achieved a tensile strength of 80.0 plus or minus 1.4 megapascals and a flexural modulus of 2.80 plus or minus 0.14 gigapascals, confirming that fire safety did not come at the expense of structural integrity. Electrical conductivity settled near 1.0 times ten to the minus four siemens per meter, a level low enough for insulation purposes yet sufficient for sensing.

That sensing capability is where the material becomes genuinely futuristic. The coating exhibited a thermal sensing response of 8.0 plus or minus 0.5 percent and an ammonia response of 12.0 plus or minus 0.9 percent, meaning its electrical resistance shifts measurably when temperature changes or when ammonia gas is present. This behavior arises from the percolating network of carbon nanotubes embedded in the epoxy. When the coating is heated, stretched, or exposed to certain vapor molecules, the tunneling gaps and contact geometry between nanotubes change, altering the conductive pathways. A coating that can report its own temperature, detect structural strain, or flag the presence of hazardous chemicals transforms a passive protective layer into an active element of a monitoring system, with obvious implications for fire early warning and structural health monitoring.

Crucially, the predictions held up in the laboratory. Prediction errors generally remained below 5 percent across the key properties, and the team reinforced confidence through repeated cross-validation and independent validation experiments on formulations the model had not seen during training. Structural analyses then explained why the optimized composition works so well. Microscopy and related characterization confirmed intercalated and exfoliated nanoclay domains dispersed through the epoxy, continuous conductive pathways formed by the carbon nanotubes, and strengthened interfaces between the epoxy matrix and both filler types. These three structural features jointly govern the multifunctional performance, and the machine learning model effectively learned to navigate the trade-offs among them without needing an explicit physical theory of each interaction.

The broader significance of the work lies in its reproducibility as a strategy rather than in any single formulation. Epoxy resins are ubiquitous in coatings, adhesives, electronics encapsulation, and composite matrices, and fire safety regulations continue to tighten across industries. Halogenated flame retardants, once standard, face increasing environmental scrutiny, pushing researchers toward nanofiller-based solutions such as clays, carbon nanotubes, graphene, and layered double hydroxides. At the same time, the literature on machine learning in polymer nanocomposites has grown rapidly, with studies applying neural networks and ensemble models to predict tribological, thermal, and dielectric properties. What distinguishes this study is the integration of the entire pipeline, from efficient experimental design through interpretable machine learning to multi-objective optimization and experimental confirmation, applied to a coating that must satisfy five demanding criteria at once.

The authors acknowledge support from Hanoi University of Industry through its Advanced Materials and Sustainable Technologies research group, and they report no competing interests. Their framework, they argue, offers a reproducible template for designing fire-safe, durable, and self-sensing epoxy nanocomposite coatings, and the approach extends naturally to other multifunctional material systems where competing property targets make exhaustive experimentation impractical. As laboratories worldwide confront similar multi-criteria design problems, the message of this research is clear: the fastest route through a vast formulation space may no longer be a systematic march through every combination, but a well-trained algorithm that knows where to look.

Subject of Research: Machine learning-assisted optimization of epoxy/MWCNT–nanoclay hybrid nanocomposite coatings for combined fire safety, mechanical performance, and self-sensing functionality

Article Title: (Machine learning-assisted optimization of multifunctional Epoxy/MWCNT–nanoclay nanocomposites for fire-safe, self-sensing coatings)

Article References: Nguyen, T. A., Dang, H. T., & Nguyen, V. H. (2026). (Machine learning-assisted optimization of multifunctional Epoxy/MWCNT–nanoclay nanocomposites for fire-safe, self-sensing coatings). Polymer Bulletin, 83(11), Article 633. https://doi.org/10.1007/s00289-026-06687-w

Image Credits: AI Generated

DOI: 10.1007/s00289-026-06687-w

Keywords: epoxy nanocomposites, multi-walled carbon nanotubes, nanoclay, flame retardancy, self-sensing coatings, machine learning, Random Forest, response surface methodology, Box–Behnken design, SHAP interpretability, structural health monitoring, polymer coatings