Machine Learning Models Accurately Predict Bacillus Sphaericus Bio-Concrete Compressive Strength

A computational evaluation of bacterially enhanced self-healing concrete has demonstrated that supervised machine learning algorithms can reliably forecast compressive strength across diverse curing periods, with tree-based ensemble methods significantly outperforming traditional neural network architectures. Civil engineering researchers Parikshit Hurukadli and B. S. Santhosh evaluated four prominent machine learning algorithms—Extreme Gradient Boosting, Random Forest, a Multilayer Perceptron, and a shallow Artificial Neural Network—to model the complex mechanical behavior of bio-concrete incorporating the bacterium Bacillus sphaericus. Developing accurate analytical formulations for bio-concrete has historically proven exceptionally difficult because the composite material couples standard Portland cement hydration chemistry with dynamic biological calcite precipitation. Across an experimental dataset of 120 structured laboratory observations, the Extreme Gradient Boosting model achieved the highest predictive fidelity on hold-out testing data, registering a coefficient of determination of 0.97 alongside a root mean square error of 2.85 megapascals and a mean absolute percentage error of 1.98 percent. These empirical findings confirm that data-driven soft computing frameworks can successfully decipher the nonlinear interplay between biological dosages and physical mix design proportions without requiring exhaustive trial-and-error laboratory batches.

The engineering challenge motivating this computational investigation stems directly from the mechanical trade-offs and chemical complexities inherent to self-healing cementitious composites. While the construction sector generates approximately 8 percent of global carbon dioxide emissions and expends substantial resources on continuous structural maintenance, bio-concrete addresses nascent cracking through autonomous microbially induced calcite precipitation. However, the incorporation of bacterial cultures, nutrient media, and chemical precursors fundamentally alters early hydration kinetics. In the experimental phase of the investigation, concrete cubes measuring 150 millimeters were cast using Ordinary Portland Cement 53 Grade conforming to IS 12269:2013, Zone II natural river sand, and 20-millimeter nominal coarse aggregate. The bacterial mixes incorporated viable Bacillus sphaericus strain ATCC 14577 introduced into the mixing water at three distinct concentration thresholds of 100,000, 1,000,000, and 10,000,000 colony-forming units per milliliter. To sustain metabolic ureolytic activity and drive calcium carbonate crystallization, the bacterial suspension was supplemented with a nutrient solution containing two grams per liter of urea and 25 millimolar calcium chloride dihydrate.

Physical laboratory testing revealed distinct, time-dependent mechanical trends that clearly distinguish conventional mixtures from bacterially modified concrete across designated curing ages of 7, 14, and 28 days under standard potable water immersion at 27 degrees Celsius. In marked contrast, bio-concrete exhibited an early-age strength penalty, registering only 9.80 megapascals at 7 days, which corresponds to roughly 74 percent of the control specimen strength. The researchers attributed this initial mechanical suppression to bacterial cells and dormant endospores interfering with early cement hydration immediately following rehydration. Over extended curing durations, however, microbial metabolic activation facilitated progressive calcite deposition within the pore network, narrowing the relative performance deficit. By day 28, bio-concrete compressive strength reached approximately 85 percent of the conventional concrete baseline, demonstrating that while bio-concrete is ill-suited for fast-track construction demanding high early strength, its ongoing biomineralization steadily closes the gap over time.

Complementary transport and surface hardness assessments highlighted additional nuances regarding how bacterial inclusion impacts matrix integrity and internal pore connectivity. Capillary water absorption measurements carried out in accordance with ASTM C1585-13 indicated that bio-concrete absorbed marginally less water at the initial 5-minute interval than conventional concrete, registering 0.027 grams compared to 0.031 grams for the control mix. This initial reduction suggests localized surface pore sealing through microbial precipitation. However, over cumulative absorption intervals of 15 and 25 minutes, bio-concrete exhibited higher water uptake, absorbing 0.062 grams and 0.074 grams respectively, compared to 0.050 grams and 0.059 grams for conventional concrete. Rebound hammer testing performed under IS 13311 Part 2 showed comparable mean surface hardness numbers between conventional concrete at 15.9 and bio-concrete at 15.6, both spanning a broad rebound range between 10 and 22. These localized surface variations and increased long-term capillary suction indicate that microbial distribution and nutrient delivery pathways can introduce microstructural heterogeneity and residual voids before complete calcification occurs.

To model these intricate mechanical behaviors computationally, the authors assembled an input feature matrix comprising five foundational parameters: curing age in days, bacterial concentration in colony-forming units per milliliter, water-cement ratio ranging from 0.45 to 0.55, nutrient concentration in grams per liter, and cement content. Compressive strength served as the sole target output variable. Before executing model training, the researchers conducted data screening using the interquartile range method to detect and filter abnormal outliers. Because the operational input parameters spanned disparate numerical magnitudes and physical dimensions, Min-Max normalization was applied to scale all features uniformly between zero and one. The complete dataset of 120 experimental observations was subsequently partitioned through stratified random sampling into an 80 percent training subset comprising 96 specimens and a 20 percent hold-out test subset containing 24 records, ensuring equal representation of all bacterial dosage levels across both evaluation partitions.

The training and hyperparameter optimization strategy relied on systematic grid searches conducted across 5-fold cross-validation folds strictly isolated within the 96-sample training set, completely reserving the hold-out testing set for final model validation. For Extreme Gradient Boosting, parameter exploration identified an optimal ensemble of 200 estimators, a maximum tree depth of 6, a learning rate of 0.05, and a subsampling ratio of 0.8. The Random Forest algorithm was configured with 300 decision trees operating under unrestricted tree depth and a minimum split requirement of two samples per internal node. The Multilayer Perceptron utilized a feedforward architecture comprising two fully connected hidden layers containing 64 and 32 neurons with Rectified Linear Unit activation, trained using the Adam optimizer at a learning rate of 0.001 with 20 percent dropout regularization to prevent overfitting across 500 epochs. The shallow Artificial Neural Network incorporated 128 hidden neurons utilizing sigmoid transfer functions, optimized rapidly through the second-order Levenberg-Marquardt backpropagation algorithm.

Comparative statistical evaluation on the unseen hold-out test set established a strict, consistent performance hierarchy across all quantitative metrics, with ensemble tree architectures markedly outperforming neural networks. Extreme Gradient Boosting delivered the most precise predictions, achieving an R-squared of 0.97, a root mean square error of 284.6 kilonewtons per square meter, and a mean absolute error of 196.3 kilonewtons per square meter. Random Forest followed closely in second position, recording an R-squared of 0.95, a root mean square error of 341.2 kilonewtons per square meter, and an MAE of 228.7 kilonewtons per square meter, while demonstrating the lowest cross-validation variance across folds at plus or minus 0.018. The Multilayer Perceptron attained third place with an R-squared of 0.93 and an RMSE of 402.8 kilonewtons per square meter, whereas the shallow Artificial Neural Network produced the largest deviations, yielding an R-squared of 0.91, an RMSE of 468.1 kilonewtons per square meter, and an MAE of 312.6 kilonewtons per square meter. Both ensemble models produced prediction intervals confirming that 90 percent of compressive strength estimates fell within plus or minus 420 kilonewtons per square meter of experimental observations.

Detailed residual examinations provided important theoretical insights into why specific algorithmic structures struggled with particular operational domains of the bio-concrete data. Analysis of error distributions confirmed that Extreme Gradient Boosting maintained a near-Gaussian, zero-centered residual profile with minimal standard deviation. The Multilayer Perceptron exhibited a platykurtic error distribution, reflecting bounded variance across intermediate curing stages where the transition from cement hydration to bacterial calcification occurs. In contrast, the shallow Artificial Neural Network exhibited a pronounced positive residual skewness of 0.61, characterized by systematic under-prediction among high-strength, high-dosage concrete specimens. The authors explained that this architectural limitation stemmed from the combination of a single hidden layer and saturating sigmoid activation functions. When elevated bacterial concentrations and extended curing ages jointly forced weighted inputs into the asymptotic plateaus of the sigmoid function, incremental parameter increases produced negligible output response. Furthermore, the absence of explicit dropout or regularization penalties in the Levenberg-Marquardt training routine allowed the network to smooth over less-frequent high-strength specimens.

Model interpretability was rigorously assessed through SHapley Additive exPlanations and permutation feature importance metrics to identify the primary physicochemical drivers governing compressive strength. Across the evaluated models, curing age and bacterial concentration emerged as the two most influential input variables, with curing age contributing 38.4 percent of the predictive weight in the XGBoost formulation, followed by bacterial concentration at 27.1 percent, water-cement ratio at 19.6 percent, and nutrient concentration contributing 5.1 percent. Partial dependence analysis extracted from the gradient-boosted trees revealed a distinct nonlinear threshold behavior regarding bacterial loading: compressive strength increased rapidly as bacterial counts rose from 100,000 to 1,000,000 colony-forming units per milliliter, but reached a clear saturation ceiling beyond 5,000,000 colony-forming units per milliliter. This saturation effect indicates that concrete pore networks contain a finite density of viable nucleation sites for calcium carbonate deposition. Adding surplus microbial biomass beyond this threshold yields negligible mechanical benefit while introducing economic inefficiencies and potential compaction voids.

Despite the high predictive accuracy demonstrated by the ensemble models, the authors explicitly emphasized several fundamental experimental and methodological constraints that govern how these findings should be interpreted. The underlying database remains restricted to 120 specimens prepared from a single bacterial strain, one Ordinary Portland Cement grade, and standardized laboratory curing tanks, meaning the trained algorithms function primarily as specialized screening tools rather than generalized universal predictors for arbitrary commercial formulations. Crucially, the investigation did not perform direct microstructural or mineralogical characterizations, such as scanning electron microscopy, energy-dispersive X-ray spectroscopy, X-ray diffraction, Fourier-transform infrared spectroscopy, or thermogravimetric analysis. Consequently, the attribution of macroscopic mechanical gains and capillary sorption changes to microbially induced calcite precipitation remains an inferred mechanism supported by external literature rather than directly observed crystallographic verification. The researchers concluded that future research must integrate comprehensive microstructural analyses, evaluate multi-strain bacterial consortia under aggressive field exposure conditions, and incorporate Bayesian hyperparameter optimization to establish robust, code-ready design protocols for durable bio-concrete infrastructure.

Subject of Research: Machine learning prediction of bacterial self-healing bio-concrete compressive strength across curing intervals

Article Title: Machine learning prediction of bacillus sphaericus bio-concrete strength: a comparative study of XGBoost, random forest, MLP, and ANN

Article References: Hurukadli, P., & Santhosh, B. S. (2026). Machine learning prediction of bacillus sphaericus bio-concrete strength: a comparative study of XGBoost, random forest, MLP, and ANN. Discover Civil Engineering, 3(1), Article 224. https://doi.org/10.1007/s44290-026-00625-5

Image Credits: AI Generated

DOI: 10.1007/s44290-026-00625-5

Keywords: bio-concrete, Bacillus sphaericus, compressive strength, machine learning, Extreme Gradient Boosting, Random Forest, Multilayer Perceptron, Artificial Neural Network, microbially induced calcite precipitation, curing age, SHAP analysis, sorptivity, rebound hammer testing, mix design optimization, concrete durability, water-cement ratio