The optimized epoxy benzoxazine composite carried 30 percent basalt fiber, 10 percent waste cotton textile fiber, and 1 percent graphene nanoplatelets, and it was cured at 160 degrees Celsius. It reached 138 MPa tensile strength and 208 MPa flexural strength. Impact strength landed at 42 kJ/m2 and Shore D hardness at 87. Storage modulus came in at 5.2 GPa, with a glass transition temperature of 132 degrees Celsius.
Why blend benzoxazine into epoxy at all
Epoxy is the default thermoset matrix for a reason. It sticks well, it carries load well, and it forgives a lot during processing. Its weak points are brittleness, limited thermal stability, and wear resistance that falls off under sliding contact.
Benzoxazine covers those gaps, then introduces a different set of problems. It polymerizes by ring-opening and releases no volatile by-products, so shrinkage during cure is low, thermal stability is high, and tribological behavior is better. Neat benzoxazine, though, is brittle and short on impact resistance. Blending the two is a way to get epoxy toughness and processability alongside benzoxazine thermal performance in one epoxy benzoxazine composite.
The hybridization is only half the story. Two fibers and one nanofiller also go into the same matrix, which means the interface between reinforcement and matrix becomes the deciding factor rather than the resin alone.
The four formulation levers and the levels tested
| Factor | Levels tested | What it is expected to move |
|---|---|---|
| Basalt fiber content | 20, 25, 30 wt% | Load bearing, stiffness, crack paths |
| Waste cotton textile fiber | 10, 15, 20 wt% | Impact energy absorption, sustainability |
| Graphene nanoplatelets | 0.5, 0.75, 1.0 wt% | Interfacial bonding, crack bridging, wear |
| Curing temperature | 140, 150, 160 degrees Celsius | Crosslink density, ring-opening extent |
Nine formulations were run under a Taguchi L9 orthogonal array, then ranked by signal-to-noise ratio using a larger-the-better criterion for tensile strength, flexural strength, impact strength, and Shore D hardness. The winning combination was 30 percent basalt fiber, 10 percent waste cotton textile fiber, 1 percent graphene nanoplatelets, and a 160 degrees Celsius cure.
What the optimized epoxy benzoxazine composite actually measured
| Property | Best sample (S7) | Weakest sample (S1) |
|---|---|---|
| Tensile strength | 138 MPa | 84 MPa |
| Flexural strength | 208 MPa | 132 MPa |
| Impact strength | 42 kJ/m2 | 22 kJ/m2 |
| Shore D hardness | 87 | 74 |
| Storage modulus | 5.2 GPa | Not reported as a single value |
| Glass transition temperature | 132 degrees Celsius | Not reported as a single value |
The spread between best and worst is the useful part here. Tensile strength moved by 64 percent and impact strength nearly doubled, and the only things that changed were reinforcement loading and cure temperature. That is the argument for running a designed experiment instead of moving one variable at a time.
The weakest sample sat at the opposite corner of the design space, with 20 percent basalt fiber, 0.5 percent graphene nanoplatelets, and a 140 degrees Celsius cure. Low reinforcement and a low cure temperature compound each other. That is the pattern you would expect if crosslink density and interfacial bonding were both limited.
Surface roughness tracks dispersion, and it bottoms out at the same sample
Atomic force microscopy gave a neat, monotonic story. Surface roughness fell from 81.2 nm on the weakest sample down to 47.6 nm on the best. Intermediate samples landed at 76.5, 71.4, 66.8, 62.3 and 59.8 nm as reinforcement loading rose. The two samples that overshot on cotton fiber and undershot on graphene came back up to 51.4 and 54.2 nm.
Roughness here is a proxy worth reading carefully. A smoother cured surface in this system goes with better fiber wetting and more even nanoparticle dispersion, and that is consistent with the mechanical ranking. It is not itself a measure of interfacial strength.
What the fracture surfaces showed
Fractography was quantified rather than just described. On the weakest sample, fiber pull-out counted 16 features per image, the void area fraction was 4.8 percent, and crack density was 12.4 cracks per square millimeter. On the improved samples those three figures fell to 8 per image, 2.3 percent, and 6.5 cracks per square millimeter, while graphene dispersion rose to 87 percent.
That is a coherent picture. Fewer voids, fewer cracks, less pull-out, better dispersion. The authors are explicit that the proposed mechanisms, meaning stress transfer and crack bridging, are physically reasonable rather than experimentally proven. They were inferred from the combined observations, never measured directly.
Curing temperature and curing agent selection: why 160 degrees Celsius won
Cure temperature ranked second in the signal-to-noise analysis for tensile strength, flexural strength, and hardness, behind basalt fiber content. Basalt fiber posted the largest delta values, at 2.93 dB for tensile, 2.76 dB for flexural, 3.45 dB for impact, and 1.63 dB for hardness. Cure temperature followed at 0.84, 0.77, 0.78, and 0.38 dB.
Anyone who formulates thermosets will recognize that ordering. Fill the matrix with enough load-bearing fiber and you set the ceiling. Then the cure schedule decides how much of that ceiling you actually reach, because it controls how far ring-opening goes and how tight the network ends up. It is the same lever a formulator pulls through curing agent selection, crosslinker level, and bake schedule in other thermoset families, polyurethane included.
The practical point is that 160 degrees Celsius was not chosen because hotter is simply better. It was the top of three levels, and the data picked the top of the range. A wider temperature window might have moved the optimum again.
Thermal stability and where the crosslinked network ends up
| Material | Onset temperature | Maximum degradation temperature | Char yield |
|---|---|---|---|
| Optimized hybrid (S7) | 335 degrees Celsius | 438 degrees Celsius | 34 percent |
| Epoxy reference range | 290 to 305 degrees Celsius | 350 to 400 degrees Celsius | 10 to 15 percent |
| Benzoxazine reference range | 318 to 325 degrees Celsius | 410 to 430 degrees Celsius | 22 to 30 percent |
The hybrid sits above both reference ranges on all three measures. Char yield of 34 percent is the number carrying the most weight, because a high char yield usually means a tightly crosslinked network that builds a protective carbonaceous layer instead of volatilizing.
One caveat matters here. Fourier-transform infrared peak intensities were normalized to an internal reference band, so they are not proportional to crosslink density. Shifts in the oxazine ring band around 810 to 830 per cm are consistent with differences in ring-opening extent. They should not be read as a quantitative conversion figure.
What the machine learning layer adds and does not
Random Forest and XGBoost models were trained on the nine formulations to predict the four mechanical responses, and XGBoost gave the better fit, with R squared between 0.9870 and 0.993.
That number deserves a cold reading. Nine data points across four input factors is an extremely small training set, and a model with that much capacity will fit nine points almost perfectly whether or not it captured the real relationships. A high R squared on this design is not evidence the model will extrapolate to a formulation outside the tested space.
What the modeling does usefully provide is a sensitivity ranking that agrees with the Taguchi analysis. Agreement between two independent methods on which factor matters most is worth more than either fit statistic.
What a formulator should take from this
- Reinforcement loading sets the ceiling in this epoxy benzoxazine composite, and cure schedule decides how much of that ceiling is reached. Rank the levers before spending experimental budget.
- The optimum sat at the highest graphene level tested, 1 percent. That is a boundary optimum, so the real optimum may sit at or above it.
- Waste cotton textile fiber helped up to 10 percent and hurt above it. Low stiffness and higher moisture uptake are the likely reasons.
- Treat FTIR band intensities as qualitative here. They were normalized to an internal reference and are not a conversion metric.
- Put cure schedule and curing agent choice into the same designed experiment as reinforcement loading, because in practice they move the same interfacial properties.
- Do not read the machine learning R squared as proof of predictive power outside the tested design space.
Limitations
The Taguchi L9 array used all eight available degrees of freedom to estimate four three-level factor effects, which leaves no independent degrees of freedom to estimate two-factor or higher interactions. Interactions were not treated as statistically significant in this design, so any interaction between reinforcement loading and cure temperature is unmeasured rather than absent.
The authors state that the proposed mechanisms, including stress transfer and crack bridging, are physically reasonable but not directly quantified and not experimentally proven. FTIR intensities were normalized to an internal reference band and are explicitly not proportional to crosslink density. The machine learning models were trained on nine formulations, so their reported R squared values describe fit to that small set and should not be read as out-of-sample accuracy. Graphene nanoplatelets were tested only up to 1 percent, so the optimum at that level is a boundary result. Wear behavior is reported as trends across 10 to 50 N load and 1 to 5 m per second sliding speed without a single headline figure given here.
Bottom line
An epoxy benzoxazine composite carrying 30 percent basalt fiber, 10 percent waste cotton textile fiber, and 1 percent graphene nanoplatelets, cured at 160 degrees Celsius, reached 138 MPa tensile, 208 MPa flexural, 42 kJ/m2 impact, Shore D 87, a 5.2 GPa storage modulus, and a Tg of 132 degrees Celsius. Basalt fiber content dominated the ranking, and cure temperature came second. The mechanisms behind those gains are inferred rather than measured, and the design had no degrees of freedom left over for interactions.
FAQ
What is an epoxy benzoxazine composite
An epoxy benzoxazine composite uses a hybrid thermoset matrix made by blending epoxy resin with benzoxazine resin, then reinforces it. Epoxy brings toughness and processability, while benzoxazine brings thermal stability and low curing shrinkage. The blend is meant to avoid the brittleness of neat benzoxazine.
What did the best formulation reach
The best formulation reached 138 MPa tensile strength, 208 MPa flexural strength, 42 kJ per square meter impact strength, and 87 Shore D hardness. Its storage modulus was 5.2 GPa and its glass transition temperature was 132 degrees Celsius. Those figures came from the sample with the highest reinforcement and the upper cure temperature.
Which factor mattered most
Basalt fiber content ranked first for all four responses, with signal-to-noise deltas of 2.93 dB for tensile strength and 3.45 dB for impact strength. Curing temperature ranked second for tensile strength, flexural strength and hardness. Graphene and cotton fiber loading mattered less across the tested range.
Why was 160 degrees Celsius the best cure temperature
It was the highest of the three levels tested, at 140, 150 and 160 degrees Celsius, and it gave the best balance of properties. Higher cure temperature is associated with more complete benzoxazine ring-opening and a tighter network. The optimum being at the boundary means a wider window might move it.
How much did the weakest sample differ
The weakest sample reached 84 MPa in tensile strength, 132 MPa in flexural strength, 22 kJ per square meter in impact strength, and 74 in Shore D hardness. That is roughly 39 percent lower tensile strength and close to half the impact strength of the best sample. The difference came from lower reinforcement and a lower cure temperature.
Did surface roughness track the mechanical results
Yes, closely. Roughness fell from 81.2 nanometers on the weakest sample down to 47.6 nanometers on the best, and the intermediate samples sat neatly in between. The two samples that overshot on cotton fiber and undershot on graphene rose again to 51.4 and 54.2 nanometers, tracking their weaker mechanical results.
What did the fracture surfaces show
The weakest sample showed 16 fiber pull-out features per image, a 4.8 percent void area fraction, and 12.4 cracks per square millimeter. Improved samples fell to 8 pull-out features, 2.3 percent voids, and 6.5 cracks per square millimeter, with graphene dispersion at 87 percent. The authors call the mechanisms inferred rather than proven.
Is the FTIR evidence quantitative
No, and this is stated directly in the report. Peak intensities were normalized to an internal reference band, so they are not proportional to crosslink density. Shifts in the oxazine ring band near 810 to 830 per centimeter are consistent with differences in the extent of ring-opening. They should not be converted into a percentage of cure, because the normalization removes that information.
How good is the machine learning model really
XGBoost reported R squared between 0.9870 and 0.993, but the training set was nine formulations. A model that flexible will fit nine points closely almost regardless of the true relationship. The useful output is the sensitivity ranking, which agrees with the Taguchi analysis, not the fit statistic.
Can the results be extrapolated beyond 1 percent graphene
Only with caution. One percent was the highest level tested, so the optimum is a boundary result rather than an interior one that the data actually located. Higher loadings were never explored, and agglomeration is a known risk once dispersion stops improving with each addition. A follow-up design should extend the range upward before anyone treats 1 percent as the true optimum.
What is the main gap in this experimental design
The L9 array spent all eight degrees of freedom on four three-level factors, leaving none for interactions. Two-factor effects such as reinforcement loading combined with cure temperature were therefore never estimated. Their absence from the analysis is a design limit, not evidence that they do not exist.