Sustainable concrete
An Empirical Model for the Relative Compressive Strength of Concrete Containing Rice Husk Ash, Palm Oil Fuel Ash and Sugarcane Bagasse Ash
Department of Civil Engineering, University of Ibadan, Ibadan, Nigeria
Published November 2026
Abstract
Rice husk ash (RHA), palm oil fuel ash (POFA) and sugarcane bagasse ash (SCBA) are abundant agro-industrial residues that can partially replace Portland cement in concrete, but designers lack a simple, validated way to estimate how much strength a given replacement will cost. This study developed and tested a literature-derived empirical model for the strength ratio R, the compressive strength of an ash concrete divided by that of the plain cement concrete made with the same materials in the same laboratory. From a database of 736 records in 37 published experimental programmes, 574 ash-concrete records from 35 programmes were each paired with their own control and used for modelling. Nine candidate models were fitted by least squares with standard errors clustered by programme and tested by leave-one-study-out validation, in which every programme in turn was predicted by a model that had never seen it. The adopted model, ln R = −(5.46 − 3.94g)x² + 0.123x ln(t/28), uses the replacement fraction x (ash mass divided by total binder mass), the curing age t in days and an indicator g that equals 1 when the source study reported milling the ash to standard fineness or beyond, as classified from the processing it described, and 0 otherwise. All three terms were statistically significant. Coding grinding in four ordered levels did not improve predictions, and neither did separate terms for the three ashes (p = 0.74) or loss on ignition. On programmes it had not seen, the model predicted R with a root-mean-square error of 0.166, against 0.205 for a constant median (bootstrap 95% interval of the reduction 0.007 to 0.067), and explained about one-third of the variance in R. The model indicates that unground ash can replace about 9% of the binder for a 5% loss of 28-day strength and that milling raises this to about 18%, or 23% if strength is assessed at 90 days. These are preliminary estimates for an average ash source; the scatter between sources remains large, so trial mixes are still required.
Keywords: agro-industrial ash; compressive strength; empirical model; grinding fineness; supplementary cementitious materials
1Introduction
Portland cement production accounts for a large share of global anthropogenic carbon dioxide emissions, and most of that share comes from the calcination of limestone rather than from fuel, so it cannot be removed by cleaner kilns alone [1]. Partial replacement of clinker with supplementary cementitious materials is one of the few measures that can reduce it at scale. In industrialised economies this route rests on coal fly ash and ground granulated blast-furnace slag, which typically replace on the order of 10 to 50% of the cement in concrete [2]. Neither material is produced in useful quantities in most of sub-Saharan Africa and South-East Asia, where the demand for new concrete is growing fastest.
Those regions do produce large quantities of agro-industrial ash. Rice husk ash (RHA) remains after husk is burnt for energy at rice mills, palm oil fuel ash (POFA) after fibre and shells are burnt in palm oil mill boilers, and sugarcane bagasse ash (SCBA) after bagasse is burnt in sugar factory boilers. The three are considered together here because they share the features that govern their use: all are siliceous combustion residues of the same tropical agro-industries, all react pozzolanically with the calcium hydroxide released by cement hydration, and all are improved by the same processing, controlled burning and grinding. Their performance depends heavily on that processing: Hwang et al. [3] report that grinding an RHA from a mean particle size of 87 µm to 12 µm raised its 28-day strength activity index from 72% to 97% of the control mortar. Pooling them also gives more programmes than any one ash offers alone, and it allows the question of whether ash type matters, once processing is accounted for, to be tested rather than assumed.
Several kinds of tool exist for estimating the strength of concrete containing a pozzolan. The oldest is the efficiency factor, or k-value, introduced by Smith [4] for fly ash, which counts each kilogram of addition as k kilograms of cement in the water–cement ratio; Papadakis and Tsimas [5] later estimated k-values for several pozzolans from their composition. Wong and Razak [6] showed that k is not a constant: it falls as the replacement level rises and grows with age, which is the behaviour a model of ash concrete must reproduce. EN 206 [7] gives k-values only for fly ash (k = 0.4), silica fume and ground granulated blast-furnace slag; any other addition must first have its suitability established, so no k-value is codified for an agro-industrial ash. The strength activity index of ASTM C618 [8] requires a mortar with 20% of its cement replaced to reach 75% of the control strength at 7 or 28 days. The index screens an ash, but it says nothing about how strength changes with replacement level or age in concrete. Statistical syntheses of the literature are the next step. Wahab et al. [9] pooled published data on agricultural ashes in a meta-regression of 28-day strength on replacement level, fineness and age (R² = 0.71) and found fineness to be the strongest correlate of strength (r = 0.69). Their response, however, is absolute strength in MPa, which mixes the effect of the ash with that of each laboratory's cement and aggregates, and their model was assessed only on the data it was fitted to.
Machine learning offers a third route. Models trained on pooled RHA data [10]–[13] predict absolute strength from mix proportions with high accuracy, but they are generally evaluated on records withheld at random, which lets mixes from one programme fall on both sides of the split; when whole sources are withheld, such accuracy can fall to about zero [14]. Mirzaei et al. [15] took the step closest to this study, predicting the 28-day strength ratio of concretes containing 18 agricultural ashes with gradient-boosted trees and reaching a cross-validated R² of about 0.80, with folds built on ash type rather than on source. Their model differs from the present one in three respects: it is a black box rather than an equation, it requires the oxide composition and specific surface area of the ash, which are often unavailable at the design stage, and it predicts 28-day strength only.
The experimental literature on the three ashes is large [16]–[52]. Because each programme uses its own cement, aggregates, specimens and curing, the results are difficult to combine. Individual studies report an optimum replacement level for their own ash, often near 10%, but those optima vary widely and cannot be transferred to another ash source. The gap is therefore specific. No published model gives, in closed form, the strength of agro-ash concrete relative to its plain counterpart as a function of the quantities a designer controls, namely the replacement level, the processing of the ash and the age at which strength is assessed, and none has been tested on experimental programmes withheld from its fitting. The relative strength is the natural target for such a model: it cancels the laboratory-specific effects of cement, aggregates and specimen type, which are common to an ash mix and its control, and it answers the designer's question directly, namely how much strength a given replacement costs.
The aim of this study was to develop and validate such a model. Its objectives were: (i) to compile, from published experimental programmes, a set of ash concretes each paired with a plain control made with the same materials; (ii) to formulate and fit candidate models of the strength ratio in terms of replacement level, grinding, curing age, ash type and loss on ignition, and to select among them by their accuracy on unseen programmes; and (iii) to derive from the adopted model the replacement levels that limit the strength loss to stated values, as a basis for preliminary mix design.
2Methodology
2.1Data collection and screening
Data were collected from published experimental programmes identified through Google Scholar and Consensus, with most searches run on 20 June 2026. The search combined the names and acronyms of the three ashes with terms for concrete, compressive strength and chemical composition. A programme was included if it reported the compressive strength of a binary concrete made with Portland cement (CEM I) and a single agro-industrial ash (RHA, POFA or SCBA) at a stated curing age, together with the 28-day strength of a plain cement control cast within the same programme. Ternary and quaternary binders, blends containing industrial by-products, concretes with aggregates other than normal-weight mineral aggregates, and steam-cured or field-cured specimens were excluded; where a programme tested several curing regimes, only the water-cured series was retained. Data were extracted from 41 programmes that met these criteria on screening, and four were removed after extraction: one duplicated an earlier study; in another, the replacement levels stated in the text and in the mix table could not be reconciled; a third used ash from date palm fronds rather than palm oil fuel ash; and the fourth made two-stage concrete with recycled coarse aggregate.
The resulting database contains 736 compressive-strength records from 37 programmes published between 2006 and 2026 [16]–[52], where a record is the mean strength of one mix at one age (Table 1). Of these, 140 records are plain controls and 596 contain ash. For every ash record, the mix proportions, the curing age, the oxide composition and loss on ignition (LOI) of the ash where reported, and a description of its processing were extracted. The 28-day control strengths range from 16.8 to 48.0 MPa, so the data describe normal-strength concrete. Figure 1 traces the records from extraction to the final modelling set: of the 596 ash records, 585 could be paired with a control at the same age, and the 11 records of the two programmes that did not describe how their ash was processed were then excluded, leaving 574 paired records from 35 programmes. All models in this paper were fitted and validated on these 574 records.
| Binder | Std. | Rec. | Repl. (%) | w/b | Age (d) | f (MPa) |
|---|---|---|---|---|---|---|
| Plain control | 37 | 140 | – | 0.23–0.70 | 1–365 | 10.4–59.3 |
| RHA | 10 | 222 | 5–30 | 0.39–0.70 | 1–365 | 6.9–71.2 |
| POFA | 14 | 211 | 1–70 | 0.23–0.95 | 3–180 | 5.3–53.4 |
| SCBA | 13 | 163 | 5–50 | 0.40–0.67 | 7–180 | 7.5–55.9 |
Std., studies; Rec., records in the full database of 736 records; Repl., replacement by mass of total binder; w/b, water to binder ratio by mass, with the ash counted as binder; f, compressive strength as reported.
2.2Response variable
The response modelled is the strength ratio
where fash is the compressive strength of an ash concrete and fctrl that of the plain control from the same programme, the same series of mixes and the same curing age. A series is a set of mixes that share a control, such as the mixes made at one water–binder ratio; where more than one control result was reported for a series and age, their mean was used. The ratio removes the effect of the cement, aggregates, workmanship and specimen geometry of each laboratory, which are common to an ash mix and its control, so that cube and cylinder results can be combined without conversion. R = 1 means that the ash concrete matches the plain concrete it replaces, and R = 0.95 means a 5% loss of strength.
2.3Explanatory variables
Four quantities were considered. The replacement fraction is defined on the total binder mass,
where mash and mc are the masses of ash and cement per cubic metre of concrete; x was computed from the reported mix proportions, so studies that described replacement as a percentage of cement are expressed on the same basis. The curing age t is in days and enters as ln(t/28), so that the age term vanishes at 28 days.
Fineness could not be used directly, because the programmes reported it in incompatible forms (Blaine surface area, median particle size, sieve residue or milling time) or not at all. The processing of each ash was therefore classified into one of four ordered levels from the description in its source paper: 1, used as received or only sieved; 2, coarse or short-duration milling; 3, standard ball milling to roughly cement fineness; and 4, prolonged or intensive milling to finer than cement. Where a paper reported a fineness measure, it was used to confirm the level, and the typical median particle size is roughly 10 to 20 µm at level 3 and below about 10 µm at level 4 [53]–[56]. The levels are therefore categories of reported processing, not measured fineness values for each ash. For the adopted model the levels were merged into a binary indicator g, equal to 1 for levels 3 and 4 (ground ash) and 0 for levels 1 and 2 (unground ash); the reason for this grouping is given in the section on model selection. Of the 574 modelling records, 273 used ash at levels 1 or 2 and 301 at levels 3 or 4. The ash type and, where reported, the LOI of the ash were tested as further variables.
2.4Model formulation
The model was formulated in logarithms, so that its predictions of R are always positive and its errors are proportional rather than absolute. Its general form is
Each term follows from a physical requirement or an observed trend. The model has no intercept because a concrete without ash is its own control: at x = 0 every term vanishes, so ln R = 0 and R = 1 exactly. The replacement enters as x² because the data show little loss of strength up to about 10% replacement and an accelerating loss beyond it (the section on the compiled data); a squared term reproduces this shape with a single coefficient, and a linear term was tested as an alternative. The age term x ln(t/28) is zero at 28 days by construction, follows the logarithmic gain of strength with age that underlies the time functions of Eurocode 2 [57], and is multiplied by x because the control gains strength too: it represents only the additional late gain of the ash concrete, which should grow with the amount of slowly reacting ash. Finally, the benefit of grinding grew with the replacement level within single programmes, so grinding enters as a modifier h of the curvature of the x² term rather than as a separate offset. Three codings of h were compared: equal steps between the four levels, a separate step for each level, and a single step between unground ash (levels 1 and 2) and ground ash (levels 3 and 4).
2.5Estimation and validation
The coefficients of each candidate model were estimated by ordinary least squares, that is, by minimising the sum of squared differences between the observed and fitted values of ln R. Records from one programme share a cement, aggregates and procedures, so their errors are correlated and conventional standard errors would overstate the precision of the coefficients. Standard errors were therefore computed with the cluster-robust (sandwich) estimator, with the programme as the cluster [58], and the significance of each coefficient was judged from these.
The models were then tested by leave-one-study-out validation. Each of the 35 programmes was removed in turn, the model was refitted to the other 34, and the R of every record in the removed programme was predicted. Every record therefore received one prediction from a model that had not seen its programme, which reproduces the situation of a designer applying the model to a new ash source and laboratory. From these withheld predictions R̂ᵢ of the n = 574 observed ratios Rᵢ, three measures were computed: the root-mean-square error, RMSE = √[Σ(R̂ᵢ − Rᵢ)²/n]; the mean absolute error, MAE = Σ|R̂ᵢ − Rᵢ|/n; and the validation R², 1 − Σ(Rᵢ − R̂ᵢ)²/Σ(Rᵢ − R̄)², the share of the variance of R about its mean R̄ that the withheld predictions explain. A baseline that predicts the median observed ratio (0.94) for every mix was used for comparison.
Whether one model was more accurate than another on new programmes was tested by a paired bootstrap. In each of 2,000 resamples, 35 programmes were drawn with replacement, the RMSE of both models was computed from the withheld predictions of the records in the drawn programmes, and the difference was recorded; the 2.5th and 97.5th percentiles of the 2,000 differences give its 95% interval, and an interval that excludes zero indicates a reliable improvement. The fitted R² is reported about the mean of ln R; because the models have no intercept, this is lower than the uncentred value printed by some software, and it describes fit, not predictive accuracy. All analyses were carried out in Python with the pandas, SciPy and statsmodels libraries.
3Results and discussion
3.1Behaviour of the compiled data
Figure 2 shows the 28-day strength ratios of the paired records against replacement level. Across the three ashes the median ratio is 1.03 at replacements up to 7.5% and 1.00 at 7.5 to 12.5%, so at these levels a typical ash concrete matches its control. Beyond 12.5% the ratio falls at an increasing rate, to 0.92 at 12.5 to 17.5%, 0.85 at 17.5 to 25%, 0.78 at 25 to 35% and 0.66 above 35%. This shape, flat at first and then falling ever more steeply, is the reason for the x² form of (3).
Two further patterns guided the choice of variables. Six programmes tested the same mix with the same ash at two or more grinding levels, giving 29 matched comparisons at 28 days, and in every one of them the more finely ground ash gave the higher ratio, by a median of 0.32. The gain grew with replacement level (Spearman ρ = 0.45, p = 0.01): in Tangchirapat et al. [49], grinding raised the ratio from 0.76 to 0.99 at 10% POFA but from 0.32 to 0.87 at 40%. And of 72 ash mixes tested at both 28 and 90 days, 55 gained on their control between those ages, the median ratio rising from 0.92 to 0.98 (Wilcoxon p < 0.001). Grinding and curing age therefore both act on the ratio, and both act more strongly at higher replacement.
3.2Model selection
Table 2 compares the candidate models. Two kinds of measure are reported, and they answer different questions: the fitted R² describes how closely a model reproduces the data it was fitted to, whereas the validation errors describe how well it predicts programmes it has not seen. Selection was based on the validation errors. Replacement level alone (M1) reduces the validation RMSE from 0.205 to 0.181, and curing age (M2) changes little. Grinding coded in four equal steps (M3) lowers the RMSE to 0.173. Giving each level its own step (M4) showed why the equal-step coding is not adequate: the step from level 1 to level 2 was 2.11 (p = 0.18), the step from level 2 to level 3 was 2.14 (p = 0.003), and the step from level 3 to level 4 was −0.42 (p = 0.66). Across programmes, then, the data resolve one contrast only, between ash that was milled to standard fineness or beyond and ash that was not. A single step between those two groups (M5) gave the lowest validation errors of all the models, with an RMSE of 0.166, an MAE of 0.131 and a validation R² of 0.33, with only three coefficients, all significant, and it was adopted.
| Model | k | R² (fit) | RMSE | MAE | R² (val.) |
|---|---|---|---|---|---|
| Baseline: median ratio | 0 | – | 0.205 | 0.159 | −0.03 |
| M1: replacement only | 1 | 0.21 | 0.181 | 0.137 | 0.20 |
| M2: M1 + curing age | 2 | 0.24 | 0.180 | 0.137 | 0.21 |
| M3: M2 + grinding, 4 equal steps | 3 | 0.40 | 0.173 | 0.133 | 0.27 |
| M4: M2 + grinding, step per level | 5 | 0.45 | 0.172 | 0.135 | 0.28 |
| M5: M2 + ground or not (adopted) | 3 | 0.44 | 0.166 | 0.131 | 0.33 |
| M6: M5 + linear term in x | 4 | 0.44 | 0.171 | 0.134 | 0.29 |
| M7: M5, curvature per ash | 5 | 0.45 | 0.179 | 0.138 | 0.22 |
| M5 on LOI subset (391) | 3 | 0.44 | 0.179 | 0.143 | 0.32 |
| M8: M5 + LOI (391) | 4 | 0.47 | 0.180 | 0.144 | 0.31 |
k, number of coefficients; R² (fit), centred R² of ln R with all records; RMSE, MAE and R² (val.) are for R in leave-one-study-out validation (574 records and 35 programmes, except the last two rows).
Three extensions of M5 were rejected. A linear term in x (M6) took a negative value (−0.15, p = 0.51) and worsened prediction, so the pooled data give no support for a gain in strength at low replacement; the medians slightly above 1 in Figure 2 are within the scatter between programmes. Giving each ash its own curvature (M7) worsened the validation RMSE to 0.179, and the three curvatures did not differ significantly (Wald test, p = 0.74). Adding LOI (M8), on the 391 records that report it, left the validation RMSE practically unchanged (0.179 to 0.180). All three extensions raised the fitted R² slightly, as any added term must, but none improved prediction of unseen programmes: the extra coefficients fitted variation peculiar to the programmes in the data rather than a transferable effect.
The gains between the best models are small, and the bootstrap shows which are reliable. M5 improved on the median baseline by 0.039 in RMSE (95% interval 0.007 to 0.067), and was more accurate for 22 of the 35 programmes. Its improvement over M2, the model without grinding, was 0.014 (95% interval −0.011 to 0.032), which is not statistically significant: grinding adds little to the prediction of a new programme taken as a whole. The grinding term was nonetheless retained, for three reasons. Its coefficient is significant (Table 3), its sign agrees with all 29 within-programme comparisons, and it changes the replacement a designer can use by a factor of about two (see Implications for mix design).
3.3The adopted model
The fitted coefficients of M5 are given in Table 3. All three are statistically significant with standard errors clustered by programme, and the signs are as expected: replacement lowers the ratio, grinding reduces that penalty, and the ratio rises with age beyond 28 days.
| Coefficient | Estimate | Clustered SE | 95% CI | p |
|---|---|---|---|---|
| β₁ (x²) | −5.459 | 1.261 | −7.93 to −2.99 | < 0.001 |
| β₂ (x ln(t/28)) | 0.123 | 0.052 | 0.02 to 0.22 | 0.017 |
| β₃ (g x²) | 3.935 | 1.334 | 1.32 to 6.55 | 0.003 |
g = 1 for ash milled to grinding level 3 or 4, 0 otherwise. Fitted to 574 paired records from 35 programmes; residual standard deviation of ln R 0.19. The estimates of β₁ and β₃ are strongly correlated (r = −0.96), so their sum, the curvature for ground ash, is better determined than either alone. Across the 35 leave-one-study-out refits, β₁ ranged from −6.24 to −4.14, β₂ from 0.106 to 0.153 and β₃ from 2.65 to 4.71.
Written out, the model is
where R is the predicted ratio of the compressive strength of the ash concrete to that of a plain CEM I concrete made with the same materials and tested at the same age; x is the ash replacement as a fraction of the total binder mass, Eq. (2), valid for 0 ≤ x ≤ 0.40; t is the curing age in days, within the 7 to 180 days of most of the data; and g = 1 for ash reported as milled to standard fineness or beyond (levels 3–4) and 0 otherwise. The predicted strength of the ash concrete follows as
where fctrl is the strength of the plain concrete at the same age. For example, for a plain concrete of 30 MPa at 28 days, 20% of ground ash gives R = exp(−1.52 × 0.04) = 0.94 and a predicted strength of about 28.2 MPa.
Milling to standard fineness removes about 70% of the curvature of the unground-ash curve (3.94 of 5.46), so that ground ash loses strength far more slowly with replacement. At 20% replacement and 28 days, (4) predicts R = 0.80 for unground ash and 0.94 for ground ash. The age term is smaller: at 20% replacement, the ratio at 90 days is about 3% higher than at 28 days. The coefficients were stable across the leave-one-study-out refits, so no single programme determines them.
The two-group coding is a simplification that errs on the safe side. Within single programmes, grinding from level 1 to level 2 raised the ratio by 0.09 to 0.32 in El-said et al. [21], and grinding from level 3 to level 4 raised it by 0.06 to 0.20 in Tangchirapat and Jaturapitakkul [48]. Across programmes these steps are hidden by other differences between laboratories, so (4) treats level 2 ash as unground and level 4 ash as ground, and it is likely to underestimate both.
The rejection of ash-specific terms deserves comment. The three ashes differ in chemistry: RHA is the most siliceous, and POFA the most variable, with only 6 of its 14 reported analyses meeting the 70% minimum for SiO₂ + Al₂O₃ + Fe₂O₃ that ASTM C618 sets for natural pozzolans [8], against 9 of 10 for RHA and 11 of 15 for SCBA. Their median ratios also differ at a given replacement level. Once grinding and curing age are taken into account, however, the remaining differences between the ashes are smaller than the differences between programmes using the same ash. This result should be read narrowly. It does not show that the three ashes perform alike; it shows that, with 10 to 13 programmes per ash, an effect of ash type could not be separated from the effect of the laboratory, whereas the effect of grinding was large enough to be detected.
3.4Validation on unseen studies
Table 4 and Figure 3 report the accuracy of (4) on programmes withheld from fitting. Overall, the RMSE is 0.166 and the MAE 0.131, against 0.205 and 0.159 for the median baseline, a reduction of about 19%. Half of all predictions fell within ±0.10 of the observed ratio and 77% within ±0.20, and the mean bias was small (−0.016). Accuracy was similar for unground and ground ash. The model was most accurate for POFA and SCBA, for which the predicted and observed ratios were well correlated (Spearman ρ = 0.75 and 0.63), and least accurate for RHA (ρ = 0.33). More than half of the RHA error comes from a single programme [52], whose ratios range from 0.31 to 1.49 across its ashes, mixes and ages, and two others [38], [44] obtained ratios about 0.18 higher than predicted.
| Subset | Records / studies | RMSE | MAE | ±0.10 | Base. RMSE |
|---|---|---|---|---|---|
| All records | 574 / 35 | 0.166 | 0.131 | 49% | 0.205 |
| RHA | 211 / 10 | 0.184 | 0.142 | 47% | 0.199 |
| POFA | 205 / 13 | 0.151 | 0.122 | 52% | 0.223 |
| SCBA | 158 / 12 | 0.160 | 0.126 | 48% | 0.189 |
| Unground ash | 273 / 22 | 0.166 | 0.134 | 45% | 0.222 |
| Ground ash | 301 / 17 | 0.167 | 0.128 | 52% | 0.189 |
| 28 days | 181 / 35 | 0.167 | 0.131 | 50% | 0.199 |
| 90 days or more | 120 / 13 | 0.179 | 0.136 | 48% | 0.202 |
errors are in units of the strength ratio R. ±0.10, share of predictions within ±0.10 of the observed ratio. Base., baseline: the median observed ratio applied to every record. Some programmes tested both unground and ground ash.
Figure 3 also shows the limits of the model. Its predictions span a narrower range than the observations: very weak and very strong ash concretes are pulled towards the middle, as expected of a model that describes the average response across many ash sources. The validation R² of 0.33 means that the model explains about one-third of the variation of R between records; the remaining two-thirds is substantial and reflects differences between ashes and laboratories that replacement level, grinding and age cannot capture. In practical terms, 80% of the observed ratios lay between 0.76 and 1.27 times the predicted value. The model is therefore useful for estimating the typical strength penalty of a replacement and for comparing options, not for predicting the strength of a particular mix with precision.
3.5Implications for mix design
Table 5 inverts (4) to give the largest replacement at which the predicted ratio stays at or above 0.95 or 0.90, that is, a strength loss of at most 5% or 10% relative to the plain concrete, and Figure 4 plots the corresponding curves. Two values are given for each case. The central value uses the fitted coefficients and applies to an average ash source. The lower value is the replacement at which the lower one-sided 95% confidence limit of the mean ratio reaches the target; it allows for the uncertainty in the coefficients and is the more cautious choice.
| Ash and age | R ≥ 0.95, central | R ≥ 0.95, lower | R ≥ 0.90, central | R ≥ 0.90, lower |
|---|---|---|---|---|
| Unground, 28 d | 9 | 8 | 13 | 11 |
| Unground, 90 d | 11 | 9 | 15 | 12 |
| Ground, 28 d | 18 | 15 | 26 | 22 |
| Ground, 90 d | 23 | 17 | 31 | 24 |
unground, grinding levels 1–2; ground, levels 3–4. Central, from the fitted coefficients; lower, from the lower one-sided 95% confidence limit of the mean predicted ratio. Values are preliminary estimates for an average ash source; an individual ash source can depart further from the mean (the section on validation).
For unground ash, a replacement of about 9% costs no more than 5% of the 28-day strength, and about 13% costs no more than 10%. Individual studies commonly report their highest strength near 10% replacement. Equation (4) has no such maximum, because the pooled data showed no gain in strength at low replacement (M6), so 10% is better read as the level up to which the loss is small than as an optimum. Grinding changes the picture substantially. Milling to standard fineness raises the 5% limit to about 18% at 28 days, and the predicted loss at 30% replacement is 13%, against 39% for unground ash. Assessing strength at 90 days adds a further 2 to 5 percentage points of replacement. Since the three ashes could not be distinguished once grinding was accounted for, these limits are offered for all three, subject to the larger uncertainty found for RHA.
An engineer would use the model in four steps: classify the ash as ground or unground from its supplier's processing; read a trial replacement from Table 5, or compute R from (4) for any x up to 0.40; estimate the strength of the ash concrete from (5), using the strength of the plain concrete it replaces; and confirm the result with a trial mix using the actual ash. The limits in Table 5 are preliminary estimates for proportioning, not design values for a particular ash, and they concern compressive strength only; workability, which unburnt carbon and fine grinding both affect, and durability must be checked separately before a mix is adopted.
4Limitations
The model inherits the limitations of the published data. Fineness was reported inconsistently, or not at all, so grinding enters as a classification of the processing each paper described rather than as a measured property, and an independent recoding would strengthen confidence in β₃. Ground ash at level 4 comes from RHA and POFA programmes only, and no SCBA programme milled its ash that far. The three ashes are unequally represented (10 to 13 programmes each), and programmes contribute unequal numbers of records; although the standard errors and the validation both respect the grouping by programme, prolific programmes still weigh more in the fit. Variables that differ between programmes, such as ash chemistry, combustion conditions, curing practice and the strength of the base concrete, could not be separated from one another, which is why LOI and ash type did not improve the predictions and why two-thirds of the variance of R remains unexplained. The evidence beyond 28 days comes from 13 programmes, and only a few extend beyond 180 days. Data above 40% replacement are sparse (13 records, 12 of them POFA), and the model underpredicted them by 0.14 on average, so it should not be used beyond 40% replacement, nor outside normal-strength concrete (28-day control strength of 16.8 to 48.0 MPa). Finally, the model predicts compressive strength only.
5Conclusion
A literature-derived empirical model of the strength of concrete containing RHA, POFA or SCBA relative to its plain counterpart was developed from 574 paired records in 35 published programmes and tested by leave-one-study-out validation.
- 1. The principal contribution is a closed-form model, R = exp[−(5.46 − 3.94g)x² + 0.123x ln(t/28)], in which all three terms are statistically significant; the predicted strength of an ash concrete is R times the strength of the plain concrete.
- 2. On programmes it had not seen, the model predicted R with an RMSE of 0.166, about 19% lower than a constant median, and explained about one-third of the variance in R; the remaining variation between ash sources is substantial.
- 3. Grinding was the only ash property whose effect the pooled data resolved; separate terms for ash type and LOI did not improve prediction of unseen programmes.
- 4. As preliminary estimates, unground ash can replace about 9% of the binder for a 5% loss of 28-day strength and ground ash about 18%, rising by 2 to 5 percentage points if strength is assessed at 90 days.
The model is a tool for preliminary proportioning and for comparing options; it does not replace trial mixes with the actual ash. Future studies would make such models more precise by reporting the fineness of the ash in a standard form, such as the median particle size, and by testing the same ash at several fineness levels and ages.
Data availability
The database was compiled as part of the first author's MSc research at the University of Ibadan. The compiled database and the analysis code are available from the corresponding author on reasonable request.
Acknowledgment
The authors thank the authors of the source studies, whose published data made this work possible.
References
- [1]Scrivener KL, John VM, Gartner EM. Eco-efficient cements: potential economically viable solutions for a low-CO₂ cement-based materials industry. Cement and Concrete Research. 2018;114:2–26.
- [2]Juenger MCG, Winnefeld F, Provis JL, Ideker JH. Advances in alternative cementitious binders. Cement and Concrete Research. 2011;41(12):1232–1243.
- [3]Hwang CL, Bui LAT, Chen CT. Effect of rice husk ash on the strength and durability characteristics of concrete. Construction and Building Materials. 2011;25(9):3768–3772.
- [4]Smith IA. The design of fly-ash concretes. Proceedings of the Institution of Civil Engineers. 1967;36(4):769–790.
- [5]Papadakis VG, Tsimas S. Supplementary cementing materials in concrete: part I: efficiency and design. Cement and Concrete Research. 2002;32(10):1525–1532.
- [6]Wong HS, Razak HA. Efficiency of calcined kaolin and silica fume as cement replacement material for strength performance. Cement and Concrete Research. 2005;35(4):696–702.
- [7]British Standards Institution. BS EN 206:2013+A2:2021. Concrete – specification, performance, production and conformity. BSI, London; 2021.
- [8]ASTM International. ASTM C618-26. Standard specification for coal ash and raw or calcined natural pozzolan for use in concrete. ASTM International, West Conshohocken, PA; 2026.
- [9]Wahab A, Danso H, Appiah-Kubi E. A systematic review of agricultural ash supplementary cementitious materials as low-carbon alternatives to ordinary Portland cement from 2005 to 2025. Discover Concrete and Cement. 2026;2:61. https://doi.org/10.1007/s44416-026-00101-4
- [10]Iftikhar B, Alih SC, Vafaei M, Elkotb MA, Shutaywi M, Javed MF, et al. Predictive modeling of compressive strength of sustainable rice husk ash concrete: ensemble learner optimization and comparison. Journal of Cleaner Production. 2022;348:131285.
- [11]Amin MN, Ahmad W, Khan K, Deifalla AF. Optimizing compressive strength prediction models for rice husk ash concrete with evolutionary machine intelligence techniques. Case Studies in Construction Materials. 2023;18:e02102.
- [12]Alyami M, Nassar RUD, Khan M, Hammad AWA, Alabduljabbar H, Nawaz R, et al. Estimating compressive strength of concrete containing rice husk ash using interpretable machine learning-based models. Case Studies in Construction Materials. 2024;20:e02901.
- [13]Al-Shamasneh AR, Kewalramani M, Mahmoodzadeh A, Alghamdi A, Alnahas J, Ghazouani N, et al. Forecasting compressive strength of concrete containing rice husk ash using various machine learning algorithms. Scientific Reports. 2025;15:23839.
- [14]Kojić N, Milošević B. Compressive strength prediction of self-compacting concrete with recycled coarse aggregate using machine learning: robust multi-split evaluation and data-leakage analysis of a stacking ensemble. Technologies. 2026;14(8):474.
- [15]Mirzaei L, Fedler CB, Ghebrab T. Machine learning for relative compressive strength of concrete incorporating agricultural bio-supplementary cementitious materials. Infrastructures. 2026;11(6):190.
- [16]Alabi SA, Mahachi J. Compressive strength of concrete containing palm oil fuel ash under different curing techniques. Materials Today: Proceedings. 2021;43:1969–1972.
- [17]Bamaga SO, Hussin MW, Ismail MA. Palm oil fuel ash: promising supplementary cementing materials. KSCE Journal of Civil Engineering. 2013;17(7):1708–1713.
- [18]Batool F, Masood A, Ali M. Characterization of sugarcane bagasse ash as pozzolan and influence on concrete properties. Arabian Journal for Science and Engineering. 2020;45(5):3891–3900.
- [19]Chandru G, Vignesh V, Saravanan R. Experimental study on partial replacement of sugarcane bagasse ash in cement. International Journal of Advanced Engineering, Management and Science. 2019;5(2):154–156.
- [20]Devdhe HN, Pote RK, Nikam PA, Kasliwal SS, Ansari US, Gaikwad AD, et al. Performance of concrete using partial replacement of cement by palm oil fuel ash (POFA): eco-friendly concrete composite. International Journal of Environmental Sciences. 2025;11(22s):1–6.
- [21]El-said A, Awad A, Ahmad M, Sabri Sabri MM, Deifalla AF, Tawfik M. The mechanical behavior of sustainable concrete using raw and processed sugarcane bagasse ash. Sustainability. 2022;14(18):11181.
- [22]Gautam A, Batra R, Singh N. A study on use of rice husk ash in concrete. Engineering Heritage Journal. 2019;3(1):1–4.
- [23]Hasan NMS, Sobuz MHR, Shaurdho NMN, Meraz MM, Datta SD, Aditto FS, et al. Eco-friendly concrete incorporating palm oil fuel ash: fresh and mechanical properties with machine learning prediction, and sustainability assessment. Heliyon. 2023;9(11):e22296.
- [24]Ismail MH, Adnan MA, Deraman R. Short-term effect on mechanical strength and water absorption of concrete containing palm oil fuel ash exposed to chloride solution. International Journal of Sustainable Construction Engineering and Technology. 2021;12(3):269–279.
- [25]Jagadesh P, Ramachandramurthy A, Murugesan R. Evaluation of mechanical properties of sugar cane bagasse ash concrete. Construction and Building Materials. 2018;176:608–617.
- [26]Jha P, Sachan AK, Singh RP. Agro-waste sugarcane bagasse ash (ScBA) as partial replacement of binder material in concrete. Materials Today: Proceedings. 2021;44:419–427.
- [27]Kartini K, Mahmud HB, Hamidah MS. Strength properties of Grade 30 rice husk ash concrete. In: Proc. 31st Conference on Our World in Concrete & Structures, Singapore. 2006.
- [28]Krishna NK, Sandeep S, Mini KM. Study on concrete with partial replacement of cement by rice husk ash. IOP Conference Series: Materials Science and Engineering. 2016;149:012109.
- [29]Kulkarni MS, Mirgal PG, Bodhale PP, Tande SN. Effect of rice husk ash on properties of concrete. Journal of Civil Engineering and Environmental Technology. 2014;1(1):26–29.
- [30]Loganayagan S, Mohan NC, Dhivyabharathi S. Sugarcane bagasse ash as alternate supplementary cementitious material in concrete. Materials Today: Proceedings. 2021;45:1004–1007.
- [31]Madandoust R, Ranjbar MM, Moghadam HA, Mousavi SY. Mechanical properties and durability assessment of rice husk ash concrete. Biosystems Engineering. 2011;110(2):144–152.
- [32]Manap N, Xuan TJ. Compressive strength and microstructure of concrete using POFA as a substitute of cement. SSRN preprint 4800856. 2024.
- [33]Mangi SA, Jamaluddin N, Wan Ibrahim MH, Abd Halid A, Abdul Awal ASM, Sohu S, et al. Utilization of sugarcane bagasse ash in concrete as partial replacement of cement. IOP Conference Series: Materials Science and Engineering. 2017;271:012001.
- [34]Memon SA, Javed U, Shah MI, Hanif A. Use of processed sugarcane bagasse ash in concrete as partial replacement of cement: mechanical and durability properties. Buildings. 2022;12(10):1769.
- [35]Oyejobi DO, Abdulkadir TS, Ahmed AT. A study of partial replacement of cement with palm oil fuel ash in concrete production. International Journal of Agricultural Technology. 2016;12(4):619–631.
- [36]Panchal JN, Challagulla SP, Siva Kishore I. Influence of palm oil fuel ash on strength properties of concrete. IOP Conference Series: Materials Science and Engineering. 2021;1197:012082.
- [37]Pone J, Ash A, Kamau J, Hyndman F. Palm oil fuel ash as a cement replacement in concrete. Modern Approaches on Material Science. 2018;1(1):MAMS.MS.ID.000102.
- [38]Ponmalar V, Abraham RA. Study on effect of natural and ground rice-husk ash concrete. KSCE Journal of Civil Engineering. 2015;19(6):1560–1565.
- [39]Priya KL, Ragupathy R. Effect of sugarcane bagasse ash on strength properties of concrete. International Journal of Research in Engineering and Technology. 2016;5(4):159–164.
- [40]Quedou PG, Wirquin E, Bokhoree C. Sustainable concrete: potency of sugarcane bagasse ash as a cementitious material in the construction industry. Case Studies in Construction Materials. 2021;14:e00545.
- [41]Ramasamy V. Compressive strength and durability properties of rice husk ash concrete. KSCE Journal of Civil Engineering. 2012;16(1):93–102.
- [42]Razali AW, Ahmad Zaini A, Darmansah NF, Ibrahim SH, Zaini N, Tamjehi SD, et al. Optimization of palm oil fuel ash (POFA) content as cement replacement based on strength and porosity. Materials Science Forum. 2020;997:3–7.
- [43]Sankeeth S, Prabhath AAN, Damruwan HGH, Herath HMST, Priyadarshana HVV, Kumara SPSNBS, et al. Optimizing mechanical properties of concrete using sugarcane bagasse ash. Results in Engineering. 2025;27:105968.
- [44]Sathe S, Kangda MZ, Dandin S. An experimental study on rice husk ash concrete. Materials Today: Proceedings. 2022. https://doi.org/10.1016/j.matpr.2022.11.366
- [45]Setayesh Gar P, Suresh N, Bindiganavile V. Sugar cane bagasse ash as a pozzolanic admixture in concrete for resistance to sustained elevated temperatures. Construction and Building Materials. 2017;153:929–936.
- [46]Shehu IA, Awal ASMA. Mechanical properties of concrete incorporating high volume palm oil fuel ash. Advanced Materials Research. 2012;599:537–540.
- [47]Siddika A, Al Mamun MA, Ali MH. Study on concrete with rice husk ash. Innovative Infrastructure Solutions. 2018;3:18.
- [48]Tangchirapat W, Jaturapitakkul C. Strength, drying shrinkage, and water permeability of concrete incorporating ground palm oil fuel ash. Cement and Concrete Composites. 2010;32:767–774.
- [49]Tangchirapat W, Saeting T, Jaturapitakkul C, Kiattikomol K, Siripanichgorn A. Use of waste ash from palm oil industry in concrete. Waste Management. 2007;27(1):81–88.
- [50]Torres-Ortega R, Saba M, Arrieta-Baldovino J. Microstructural evolution and mechanical performance of concrete incorporating palm oil fuel ash as a partial cement replacement. Recycling. 2026;11:139.
- [51]Ullah MF, Tang H, Ullah A, Toth Z, Ahmad M, Alzlfawi A. Mechanical and environmental performance of sugarcane bagasse ash from Khyber Pakhtunkhwa in sustainable concrete. Scientific Reports. 2025;15:24571.
- [52]Zerbino R, Giaccio G, Isaia GC. Concrete incorporating rice-husk ash without processing. Construction and Building Materials. 2011;25(1):371–378.
- [53]Bui DD, Hu J, Stroeven P. Particle size effect on the strength of rice husk ash blended gap-graded Portland cement concrete. Cement and Concrete Composites. 2005;27(3):357–366.
- [54]Rukzon S, Chindaprasirt P, Mahachai R. Effect of grinding on chemical and physical properties of rice husk ash. International Journal of Minerals, Metallurgy and Materials. 2009;16(2):242–247.
- [55]Cordeiro GC, Toledo Filho RD, Tavares LM, Fairbairn EMR. Ultrafine grinding of sugar cane bagasse ash for application as pozzolanic admixture in concrete. Cement and Concrete Research. 2009;39(2):110–115.
- [56]Tangchirapat W, Jaturapitakkul C, Chindaprasirt P. Use of palm oil fuel ash as a supplementary cementitious material for producing high-strength concrete. Construction and Building Materials. 2009;23(7):2641–2646.
- [57]British Standards Institution. BS EN 1992-1-1:2004. Eurocode 2: design of concrete structures – part 1-1: general rules and rules for buildings. BSI, London; 2004.
- [58]Cameron AC, Miller DL. A practitioner's guide to cluster-robust inference. Journal of Human Resources. 2015;50(2):317–372.