AI for concrete mix design

Know the strength
before you pour.

Concrete tells you how strong it is twenty-eight days after it is already holding up a building. PrePour reads the mix and tells you on day zero.

Every floor is poured on trust.

A slab is cast, the formwork comes off, the next floor goes up. The cubes taken from that pour are crushed in a lab weeks later. Construction does not wait for the answer.

By the time a cube fails, five more floors are standing on it.

Indian and international practice (IS 456, IS 516, ASTM C39) judge concrete by its 28-day compressive strength. A low result means coring, load tests, jacketing or demolition. To avoid that risk, mixes carry extra cement, and cement is the expensive, carbon-heavy part.

28daysstandard age at which strength is judged, IS 456 cl. 6.2

~8%of global CO₂ comes from cement production (Chatham House, 2018)

8inputsis all a mix sheet needs to predict strength with an ANN

What this tool sets out to do

Predictcompressive strength in MPa from seven mix constituents and the age of the concrete.
Boundevery forecast with an uncertainty band, so it is never read as a single magic number.
Interpretthe result as an engineer would: IS 10262 target strength, IS 456 grade and durability limits, column capacity.
Decarbonisesearch for leaner, lower-carbon mixes that keep the predicted strength.

1030 real mixes, crushed in a lab.

The model learns from the Yeh concrete dataset (Cement and Concrete Research, 1998; UCI repository #165): 1030 laboratory specimens, each with its full mix and a measured compressive strength. The silos beside the tower hold the average mix at true scale.

Input	Unit	Range	Distribution

Output: compressive strength, MPa. Age enters on a log scale, because strength gain is close to logarithmic in time.

Five networks, one honest answer.

Each network maps 8 standardised inputs through 48 and 24 tanh neurons to a strength. Five are trained with different seeds; their mean is the forecast and their disagreement widens the band. Tested on 206 mixes it never saw.

Predicted vs measured206 held-out mixes, MPa

What drives strengthPermutation importance, share of total

Reading the result like an engineer

Strength developmentModel forecast by age, with uncertainty band

Sensitivity at 28 daysChange in MPa if each constituent moves ±10 % of its range

Durability check

IS 456 Table 5

Axial column capacity

IS 456 cl. 39.3, 2 % Fe500 steel, 5 m × 5 m bays, 13 kN/m² per floor, factor 1.5

Interpretation

Same strength. Less cement.

The model is fast enough to try thousands of mixes in a second. It swaps cement for slag and fly ash, trims water, keeps the batch at one cubic metre, and keeps only mixes that hold your predicted 28-day strength and your exposure limits.

CandidatesPredicted 28-day strength vs embodied CO₂e

rejectedmeets strength and limitsbest found

Pour with evidence.

Where it stops

Trained on one laboratory dataset. Aggregate type, size, cement brand, air content and curing temperature are not inputs.
Inputs are held to the range of the training data. Outside it, the band is not trustworthy and the page says so.
A forecast is a screening tool for mix design. Acceptance still comes from IS 516 cube tests on site.
Column and floor counts are a preliminary axial check, not a structural design.

How it was built

Data: Yeh, I-C. (1998). Modeling of strength of high-performance concrete using artificial neural networks. Cement and Concrete Research 28(12).
Model: scikit-learn MLP ensemble, trained offline (model/train.py), weights run in your browser.
CO₂ factors are typical cradle-to-gate values; use your supplier's EPDs for real decisions.

Back to the lab