⚡ electrical engineering

Optimum crumb rubber content for fly ash and stone dust stabilised subgrade: a field study

This field study identifies 2% crumb rubber as the optimal dosage for fly ash and stone dust-stabilized subgrade, demonstrating that while laboratory and field tests confirm a performance peak at this level, a high-performing synthetic machine learning model fails to capture this physical reality, highlighting the limitations of internal validation metrics in predicting actual material behavior.

Original authors: M Harishkumar, R Samundeswari, Arunraj E, Godson caleb J, Vincent Sam Jebadurai S, G Hemalatha, C Mightraj

Published 2026-10-22

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: M Harishkumar, R Samundeswari, Arunraj E, Godson caleb J, Vincent Sam Jebadurai S, G Hemalatha, C Mightraj

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The ground beneath our roads, railways, and buildings is rarely just dirt. It is a complex mixture of soil, water, and air that must be strong enough to hold up heavy traffic without sinking or cracking. When the native soil is too soft, engineers must improve it, a process known as stabilization. Traditionally, this has meant digging up vast amounts of rock and gravel to mix into the earth, a practice that consumes natural resources and creates a massive demand for quarried materials. In recent years, the search for a better way has turned toward waste. Industries produce huge quantities of byproducts like fly ash, a fine powder left over from burning coal, and stone dust, the tiny particles created when rocks are crushed. Meanwhile, cities struggle with mountains of discarded tires. The idea of mixing these waste materials into the soil to make it stronger is appealing: it solves two problems at once by strengthening the ground and diverting trash from landfills. However, simply throwing waste into the soil does not guarantee a better road. Different materials behave in different ways; some harden the soil like cement, while others act like rubber, making it more flexible but potentially weaker. Finding the perfect balance is difficult because adding too much of one thing can undo the benefits of the rest.

A team of researchers in India set out to solve this puzzle for a specific type of soil found in their region, known as laterite. They wanted to find the exact amount of crumb rubber—tiny pieces of shredded tire—that could be mixed with fly ash and stone dust to create the strongest possible road foundation. Their approach was unique because they did not rely solely on computer models or laboratory guesses. Instead, they built six actual sections of road in the field, each with a different amount of rubber mixed in, ranging from none at all to five percent of the total dry weight. They then tested these sections right where they were built, using two completely different methods to measure how well the ground could support weight. One method involved pressing a heavy plate into the ground to see how much it sank, while the other measured how much force was needed to push a cone into the soil. These tests are mechanically independent, meaning they use different equipment and measure different physical properties, so if both tests pointed to the same answer, the result would be highly reliable.

The researchers discovered a clear peak in performance. When they added a small amount of rubber, specifically two percent by weight, the ground became significantly stronger. Compared to a mixture with no rubber, the two-percent mix could support about ten percent more weight and was nearly twelve percent stiffer. This improvement happened because the tiny rubber particles filled the empty spaces between the soil grains, helping to distribute the pressure from vehicles more evenly without breaking the rigid structure of the soil. However, the story changed when they added more. As the rubber content increased beyond two percent, the strength of the ground began to drop steadily. By the time they reached five percent rubber, the ground was actually weaker than the mixture with no rubber at all, supporting five percent less weight. The rubber had become so abundant that it started to carry the load itself, acting like a soft cushion rather than a strong skeleton, which caused the soil grains to lose contact with one another. This confirmed that there is a specific, optimal amount of rubber for this type of soil, and adding more does not mean better performance; in fact, it makes the road foundation worse.

Part of this study was also a critical look at how engineers use computer models to predict these results. The researchers built a sophisticated artificial intelligence model trained on a large set of synthetic data—numbers generated by a computer to simulate thousands of possible mixtures. This model was very good at predicting results within its own generated data, showing high accuracy scores that would usually convince a researcher it was ready for use. However, when the team tested this model against their real-world field results, it failed completely. The computer model predicted that adding rubber would simply make the ground weaker in a straight line, missing the peak performance at two percent entirely. It also overestimated the strength of the ground by a factor of eight. The reason for this failure was not the computer model itself, but the data it was trained on. The synthetic data had been created using numbers from previous reports that contained hidden errors, such as incorrect units of measurement and values that were calculated rather than actually measured. Because the computer learned from these flawed numbers, it learned the wrong patterns. This finding serves as a powerful warning: a computer model can look perfect on paper and still be completely wrong if the data feeding it is not physically accurate.

The study concludes that for this specific combination of materials, the best approach is to use a modest amount of rubber, around two percent, mixed with twenty percent fly ash and twenty percent stone dust. This recipe diverts a significant amount of waste from landfills while actually improving the road. Adding more rubber to divert even more waste would be counterproductive, resulting in a weaker road that fails to meet engineering standards. The research also highlights the importance of checking the raw numbers behind any data set. By re-examining the original records from their own previous work, the team found misaligned data columns and incorrect units that had silently corrupted their synthetic database. Their work demonstrates that while using waste to build infrastructure is a promising path toward sustainability, it requires careful, real-world testing to find the precise limits where waste becomes a benefit rather than a liability. The ground beneath our feet is unforgiving; it responds to the exact recipe we give it, and in the case of rubber-stabilized soil, the sweet spot is narrow, precise, and easily missed if one relies only on theory.

Technical Summary: Optimum Crumb Rubber Content for Fly Ash and Stone Dust Stabilised Subgrade

Problem Statement
Pavement foundations require significant volumes of quarried granular material. While utilizing industrial and post-consumer waste (fly ash, stone dust, and crumb rubber) offers a route to reducing this demand, the mechanical behavior of rubber-amended soils is complex. Unlike pozzolanic fly ash or angular stone dust, crumb rubber is an elastomeric inclusion that reduces unit weight and increases compressibility. Consequently, rubber content is expected to exhibit an "interior optimum"—a dosage where bearing capacity improves due to void filling and stress redistribution, followed by a decline as the elastomeric phase becomes continuous and disrupts the soil skeleton. However, the specific transition dosage depends on soil gradation, rubber particle size, and compaction energy, making it difficult to predict from constituent properties alone.

Furthermore, mix design for waste stabilization is increasingly relying on data-driven surrogate models trained on synthetic corpora. A critical gap exists in validating these models: internal cross-validation metrics (e.g., $R^2$ , RMSE) often measure self-consistency with the synthetic generator rather than physical fidelity to the real material system. This study addresses the lack of field-verified optimum dosages for rubber-stabilized subgrades and the potential for synthetic models to mask physical invalidity.

Methodology
The research employed a two-stage experimental programme:

Laboratory Screening: Triplicate specimens of a lateritic subgrade were tested to establish a base mix. Unconfined compressive strength (UCS) was measured for varying combinations of stone dust (0–30%) and fly ash (0–40%). The combination of 20% stone dust and 20% fly ash was selected as the base mix due to its high mean strength (1.122 MPa) and low coefficient of variation (2.97%).
Field Construction and Testing: Six field sections were constructed on the selected base mix, differing only in crumb rubber content (0%, 1%, 2%, 3%, 4%, and 5% by dry weight). Each section was characterized in situ using two mechanically independent tests:
- California Bearing Ratio (CBR): Following IS 2720 (Part 16), measuring load-penetration response.
- Plate Load Test: Following IS 1888 using a 300 mm diameter plate to determine the modulus of subgrade reaction ( $k$ ).
Data Verification Protocol: Before analysis, all reported field values underwent a rigorous verification protocol. This included recomputing every reported quantity from primary load-penetration and load-settlement curves, checking dimensional consistency of units, and testing for circularity (whether derived values were merely correlations of other reported values).
Surrogate Model Audit: A stacked ensemble model (combining Ridge, Random Forest, XGBoost, LightGBM, and MLP with a Gaussian Process residual corrector) was trained on an 800-record synthetic corpus describing the same material system. This model was then evaluated at the specific field mix design parameters to test whether the synthetic corpus, and the model trained upon it, could reproduce the physical phenomenon of an interior optimum.

Key Results

Field Performance: Both CBR and plate load tests resolved an identical interior optimum at 2% crumb rubber. At this dosage, the CBR increased by 10.9% and the modulus of subgrade reaction by 11.9% relative to the rubber-free base mix. Beyond 2%, both metrics declined monotonically. At 5% rubber, the CBR fell 5.1% below the base mix, confirming that excess rubber is detrimental.
Surrogate and Corpus Discrepancy: The surrogate model, which achieved high nested cross-validated $R^2$ scores (0.934 for UCS, 0.939 for CBR, 0.888 for resilient modulus) on the synthetic corpus, failed to reproduce the measured field behavior when evaluated at the field mix design. It predicted a monotonic decline in performance across the 0–5% rubber range, resolving no interior optimum. Furthermore, it overpredicted the field CBR by a factor of 8.1. The study concludes that the deficiency lies in the synthetic corpus itself, which was generated from previously reported values containing errors, rather than in the model's architecture or fitting capability.
Corpus Provenance Analysis: The discrepancy was traced to the synthetic corpus's generation source. The corpus reproduced previously reported compressive strength and resilient modulus values (which contained unit mislabeling and calculation errors) within 4%, but diverged significantly from the bearing ratio. This indicated the corpus was generated from flawed reported values rather than independent physical measurements, causing the model to inherit these physical invalidities.
Verification Findings: The data verification protocol identified four specific errors in the source records used for the study:
1. A dimensionally incorrect unit for the modulus of subgrade reaction ( $kN/cm^2$ instead of $kN/cm^3$ ).
2. A misaligned load column in the Mix C CBR data.
3. A plate load result for Mix A that could not be derived from its own tabulated data.
4. A resilient modulus that was a direct correlation of the CBR (via IRC:37) rather than an independent measurement.

Significance and Claims
The paper asserts that its primary contribution is the experimental establishment of an interior optimum for crumb rubber in a specific fly ash and stone dust stabilised lateritic subgrade, verified by two independent field tests. This finding challenges the assumption that waste addition is always beneficial, demonstrating that sustainability goals (waste diversion) must be bounded by engineering performance limits.

Secondly, the study provides a methodological critique of surrogate modeling in geotechnical engineering. It demonstrates that high internal validation scores on a synthetic corpus are not diagnostic of physical validity. A model can be statistically robust yet physically wrong if the training data (the corpus) inherits errors or lacks critical physical features (like the interior optimum). The authors argue that validation must include "held-out" physical measurements that the generator never saw, specifically checking for features like the position of an optimum or the sign of a gradient.

Finally, the paper emphasizes the necessity of reproducibility protocols. It claims that simple checks—recomputing derived values from primary curves and verifying dimensional consistency—are inexpensive, general, and essential to prevent the silent propagation of errors into future datasets and models. The study concludes that for this specific material system, crumb rubber should be dosed at approximately 2% to maximize engineering benefit, while higher dosages compromise structural integrity despite increased waste diversion.

More like this