strength_predictions.js, a precomputed response surface generated from the trained GPR model. It works directly from the local file without running a Python server.
- Inputs follow the paper's five binder oxides: CaO, SiO2, Al2O3, MgO, and Na2O, plus water, aggregate, and age variables.
- The three ternary panels match Fig. 4: CaO-Al2O3-SiO2, CaO-Na2O-SiO2, and CaO-MgO-SiO2.
- All three maps use the saved strength GPR output table; the selected marker is normalised within each ternary projection.
Calculation workflow
Show calculation workflow
Precursor reaction lane
Activator and binder lane
Interpretation guide
Can the ternary colour maps guide mix design?
Yes. The coloured points come from saved GPR predictions for the selected concrete mix scale, so darker regions indicate candidate compositions with higher predicted strength in the precomputed design space.
Why can the selected point appear stronger in one map than another?
Each ternary plot is a different projection of the same higher-dimensional mix. Plot A normalises CaO-Al2O3-SiO2, plot B normalises CaO-Na2O-SiO2, and plot C normalises CaO-MgO-SiO2. A selected mix can sit near a high-strength cluster in one projection while appearing less central in another because the hidden variables are different.
Which result should be treated as the prediction?
The numeric predicted strength and uncertainty above the maps are the prediction for the current full input set. The ternary maps are visual design guidance, not separate three-variable models.
What reaction assumption is used for precursor dissolution?
The current model simplifies precursor reaction as congruent dissolution: each precursor releases its oxide components in proportion to its bulk oxide composition and degree of reaction. It does not explicitly model selective dissolution, where different elements may dissolve at different rates.
How should the optimisation result be interpreted?
There are two optimisation modes. Optimise activator / water keeps the selected precursor proportions fixed and searches practical NaOH, sodium silicate, and optional additional-water dosages. Optimise precursor proportions also varies the precursor 1 / precursor 2 ratio before suggesting a practical starting mix. Both modes compare candidates near the C-(N)-A-S-H and N-A-S-H design zones, then select the stronger reachable saved GPR prediction under the chosen binder, fine aggregate, and coarse aggregate condition. If target water/binder is fixed, additional water is calculated to match that ratio as closely as possible. The optimised marker is a suggested reference point, not a guaranteed optimum, and should be assessed with the predicted strength, uncertainty, and distance from the target chemistry zone.
Model Basis And References
Show model basis and references
This prototype builds on published work coupling machine learning, thermodynamic modelling, ternary composition maps, and inverse design concepts for alkali-activated and cementitious materials.
-
Coupling machine learning with thermodynamic modelling to develop a composition-property model for alkali-activated materials
Ke, X. and Duan, Y. (2021) Composites Part B: Engineering, 216. doi: 10.1016/j.compositesb.2021.108842. -
A Bayesian machine learning approach for inverse prediction of high-performance concrete ingredients with targeted performance
Ke, X. and Duan, Y. (2021) Construction and Building Materials, 270. doi: 10.1016/j.conbuildmat.2020.121424.
The current webtool uses a precomputed GPR response surface and ternary design maps to guide practical starting points for alkali-activated concrete mix design.