Underwater concrete repair has always been a battle against physics. When a repair grout is dropped into a dam gallery, a flooded spillway, or a submerged foundation, the surrounding water immediately begins to strip away its finest particles, diluting the mix and scattering the very material meant to restore structural integrity. Now, a research team led by Kailong Lu and Xiao Sun has combined an old biological trick with cutting-edge artificial intelligence to tackle this problem, reporting in Case Studies in Construction Materials a cementitious underwater repair grout that is designed not by trial and error, but by a machine-learning-driven optimization framework that explicitly incorporates microbes, urea, and calcium salts into its recipe.
The centerpiece of the work is a cementitious underwater repair grout, or CURG, that harnesses microbially induced carbonate precipitation, commonly known as MICP. In this natural process, ureolytic bacteria such as Bacillus pasteurii hydrolyze urea, generating carbonate ions that react with dissolved calcium to precipitate calcium carbonate in place. Inside a cementitious matrix, these mineral precipitates can plug pores, seal microcracks, and strengthen the fragile interfacial zones that normally limit the durability of underwater repairs. The researchers immobilized the bacteria within 100-mesh diatomite, a porous silica-based carrier that shields the microorganisms from the harsh alkaline environment of hydrating cement while preserving their ureolytic activity. Alongside the bacteria, the grout contained urea as a substrate, yeast extract as a nutrient source, and calcium chloride as the calcium reservoir.
But introducing living systems and their chemical feedstocks into a cement grout creates complications. The researchers used hydroxypropyl methylcellulose, or HPMC, a cellulose-derived polymer with a viscosity of 98,000 cP, to give the fresh grout its anti-washout backbone. HPMC works by increasing the viscosity and structural cohesion of the mix, but this same thickening reduces fluidity and interferes with the dispersion efficiency of polycarboxylate superplasticizers. Add the biological components on top of that, and the result is a tangled web of competing interactions. The team’s own screening experiments showed that HPMC had to be included at a minimum of 0.5 percent by mass to achieve a Plunge-test mass loss ratio of just 2.5 percent, providing the baseline underwater stability that any viable repair grout must possess.
To understand how each MICP-related ingredient affected the grout, the team conducted systematic single-factor experiments across fourteen representative mixtures. The Plunge test, which drops a screen-borne cylinder of fresh grout through a column of water three times, quantified anti-dispersion performance. Increasing urea concentrations caused the mass loss ratio to climb steadily, with the worst performance at 0.18 mol/L, suggesting that excessive urea disrupts the HPMC-stabilized flocculated network holding the grout together. Calcium chloride produced a more complex, nonmonotonic response, with the highest mass loss occurring at 0.12 mol/L. Bacterial concentration, ranging from 1×10⁴ to 1×10⁸ cells per milliliter, showed only a limited influence on underwater stability, though low concentrations caused a slight uptick in material loss.
The mechanical picture was equally nuanced. Compressive strength was tested at 3, 7, and 28 days on 70.7 mm cubic specimens under both above-water and underwater casting conditions. At early ages, both urea and bacteria reduced compressive strength, likely because their presence initially interfered with normal cement hydration. But as curing progressed, strength in bacteria-containing mixtures recovered and even surpassed that of bacteria-free counterparts, pointing toward a delayed densification mechanism. Scanning electron microscopy confirmed this: bacteria-containing specimens displayed abundant spherical mineral precipitates interwoven with rod-like, platelet-like, and rosette-like microcrystals, with some precipitates visibly occupying pore spaces that remained empty in bacteria-free controls. X-ray diffraction revealed stronger calcite reflections near 2θ = 29.4° in bacteria-containing samples, indicating enhanced carbonate precipitation. Mercury intrusion porosimetry showed a lower cumulative intrusion volume and a shift of the dominant pore-size peak toward smaller diameters, confirming genuine pore refinement rather than surface-level change.
Recognizing that these nonlinear, competing effects could never be untangled by conventional single-factor analysis alone, the researchers assembled a comprehensive experimental dataset comprising 175 compressive strength measurements, 162 splitting tensile strength measurements, and 65 fluidity measurements. Input variables included the water-cement ratio, diatomite content, HPMC content, water reducer dosage, yeast extract concentration, urea concentration, calcium chloride concentration, bacterial concentration, curing age, casting condition, and hydration time. Six regression algorithms were then compared: back-propagation neural networks, support vector regression, random forest, AdaBoost, XGBoost, and Light Gradient Boosting Machine (LGBM). All were trained with Bayesian optimization of their hyperparameters, using a Gaussian process surrogate with expected improvement as the acquisition function, totaling 40 candidate evaluations per model per target property.
XGBoost emerged as the clear winner across all three targets. For compressive strength, it achieved a testing-set coefficient of determination of 0.989, with a root-mean-square error of just 0.946 MPa and a mean absolute percentage error of 5.982 percent. For splitting tensile strength, its testing R² reached 0.965 with an RMSE of 0.036 MPa. For fluidity, it delivered a testing R² of 0.963. LGBM followed closely behind, while support vector regression notably underperformed on unseen data despite excellent training fit, a hallmark of weak generalization on small, nonlinear material datasets. Five-fold cross-validation confirmed that XGBoost’s performance was stable across different data partitions, with narrow interquartile ranges in fold-wise R² values and no evidence of severe overfitting.
To open the black box, the team applied SHapley Additive exPlanations, or SHAP, analysis to the winning XGBoost models. For mechanical properties, curing age and HPMC content dominated the feature-importance rankings, with higher curing ages driving positive contributions to strength predictions and higher water-cement ratios exerting negative effects. For fluidity, HPMC was the single most influential variable, consistently associated with reduced flowability due to its thickening action, while higher water-cement ratios pushed fluidity upward. The MICP-related components contributed measurably to mechanical-property predictions, though their individual rankings varied across models, reinforcing the idea that their effects are real but entangled with hydration dynamics and polymer flocculation.
The final step married prediction to design. The researchers coupled the BO-optimized XGBoost models with NSGA-II, the Non-dominated Sorting Genetic Algorithm II, a multi-objective evolutionary method that uses fast non-dominated sorting, crowding distance, and elitist selection to map a Pareto front of trade-off solutions. Eight mix proportion variables served as decision variables, constrained within experimentally justified ranges: water-cement ratio between 0.45 and 0.55, diatomite between 0 and 10 percent, HPMC between 0.5 and 0.7 percent, water reducer between 0.05 and 0.15 percent, yeast extract between 1 and 20 g/L, urea and calcium chloride each between 0.06 and 0.18 mol/L, and bacterial concentration between 1×10⁴ and 1×10⁸ cells/mL. With a population of 40 and 60 generations, the algorithm produced 40 Pareto-optimal grout formulations balancing compressive strength, splitting tensile strength, and initial fluidity at 28 days under underwater casting conditions.
Experimental validation of two representative Pareto schemes confirmed the framework’s practical value. Scheme A, a strength-priority design with a water-cement ratio of 0.47, 2.2 percent diatomite, 0.55 percent HPMC, 8.08 g/L yeast extract, and 1×10⁷ cells/mL bacteria, achieved a measured 28-day compressive strength of 20.87 MPa against a prediction of 19.63 MPa, a relative error of just 5.9 percent. Scheme B, a workability-priority design with 4.4 percent diatomite and 1×10⁸ cells/mL bacteria, delivered an initial fluidity of 225.5 mm against a predicted 220.5 mm, an error of 2.2 percent. Splitting tensile strength predictions showed slightly larger relative errors of 9.5 and 8.6 percent, still acceptable given the small absolute magnitudes and the inherent sensitivity of tensile behavior to material heterogeneity.
What makes this study notable beyond its immediate results is the explicit incorporation of biological and MICP-specific variables into data-driven mix design, a territory that has largely been reserved for conventional concrete parameters like aggregate content, cement replacement, and admixture dosage. Prior machine-learning-assisted mix design studies have optimized strength, durability, cost, and carbon footprint for ordinary and recycled-aggregate concretes, but rarely have they treated living bacteria and their chemical substrates as design variables with quantified trade-offs. By embedding yeast extract, urea, calcium chloride, and bacterial concentration directly into the optimization space alongside HPMC and water-cement ratio, the researchers have demonstrated that bio-augmented cementitious materials can be rationally engineered rather than discovered serendipitously.
The approach also highlights a broader trend in construction materials science: the shift from empirical, single-factor experimentation toward surrogate-model-based multi-objective optimization, where machine learning models serve as fast approximations of expensive physical tests and evolutionary algorithms navigate the resulting design space efficiently. For small experimental datasets typical of laboratory materials research, Bayesian hyperparameter tuning offers a principled way to extract maximum predictive power from limited data, and the combination with NSGA-II provides a structured way to surface trade-offs that engineers must ultimately resolve based on project priorities.
The authors acknowledge limitations that temper immediate field deployment. The dataset remains modest, particularly for fluidity prediction, and all experiments were conducted under controlled laboratory conditions with defined water depths and curing temperatures. Real underwater environments introduce currents, temperature gradients, suspended sediments, and variable hydrostatic pressure that could shift the delicate balance between anti-dispersion, fluidity, and mechanical development. No independent external validation dataset has yet been used to test the models’ generalizability beyond the investigated ranges. Nevertheless, the study establishes a replicable template: microbially enhanced underwater repair grouts, long viewed as laboratory curiosities, can now be systematically optimized using a framework that any materials laboratory with modest computational resources could adopt. As infrastructure ages worldwide and underwater repair demands grow, the marriage of MICP biology and machine learning may well become a standard tool in the engineer’s arsenal.
Subject of Research: Development and multi-objective mix design optimization of a microbially enhanced cementitious underwater repair grout using MICP-related components and a BO-XGBoost-NSGA-II machine learning framework
Subject of Research: Technology and Engineering
Article Title: Data-driven multi-objective mix design optimization of cementitious underwater repair grout incorporating MICP-related components
Article References: Lu, K., Sun, X., Wang, X., Chen, X., & Ma, H. (2026). Data-driven multi-objective mix design optimization of cementitious underwater repair grout incorporating MICP-related components. Case Studies in Construction Materials, 25, Article e06468. https://doi.org/10.1016/j.cscm.2026.e06468
Image Credits: AI Generated
DOI: 10.1016/j.cscm.2026.e06468
Keywords: underwater repair grout, MICP, Bacillus pasteurii, XGBoost, NSGA-II, Bayesian optimization, SHAP analysis, anti-dispersion, multi-objective optimization, cementitious materials, machine learning, calcium carbonate precipitation
Cite Scienmag News
APA
MLA
Chicago
Blake Davidson. (September 4, 2026). Optimizing MICP-enhanced underwater repair grout mix designs with data-driven methods. Scienmag. https://scienmag.com/optimizing-micp-enhanced-underwater-repair-grout-mix-designs-with-data-driven-methods/
Blake Davidson. “Optimizing MICP-enhanced underwater repair grout mix designs with data-driven methods.” Scienmag, 4 September 2026, https://scienmag.com/optimizing-micp-enhanced-underwater-repair-grout-mix-designs-with-data-driven-methods/. Accessed 4 September 2026.
Blake Davidson. “Optimizing MICP-enhanced underwater repair grout mix designs with data-driven methods.” Scienmag. September 4, 2026. https://scienmag.com/optimizing-micp-enhanced-underwater-repair-grout-mix-designs-with-data-driven-methods/
Copy citation
Download RIS
Tags: AI-driven design of repair compositesartificial intelligence for construction material optimizationbio-cementation for underwater repairsbio-inspired underwater grout designcalcium carbonate precipitation in concrete repaircalcium carbonate precipitation in underwater repaircalcium salts and urea in cementitious mixesdata-driven construction material formulationdata-driven optimization of cementitious groutimproving durability of underwater concrete repairsinnovative underwater grout formulationsmachine learning in construction materialsMICP-enhanced grout mix designmicrobial concrete reinforcementmicrobial incorporation in repair groutsmicrobial-based concrete repair methodsmicrobial-enhanced underwater repair grout formulationmicrobially induced calcium carbonate precipitationMicrobially induced carbonate precipitation in underwater concrete repairnatural processes in civil engineering repairporous silica carriers for bacteria immobilizationUnderwater concrete repair optimizationunderwater structural repair technologies
