Designing new concrete mixtures using optimization

No more endless trial-and-error. Using simple regression and optimization tools, we helped our client design new concrete mixtures that are more environmentally friendly, cheaper, and just as strong as hand-crafted ones. And all of this in just a fraction of the time!

Context

Our client is specialized in the development sustainable concrete mixtures by using waste product from related industries. One of the key challenges is to find appropriate quantities for each component in the concrete mix. Traditionally, this has been a lengthy trial-and-error driven by intuition.

Given the time and cost involved in preparing, curing, and testing samples, our client turned to us for a more systematic and efficient solution. Together, we have developed a method that identifies the optimal concrete recipe based on a set of requirements and a small number of example mixtures.

Approach

One of the biggest challenges in predicting concrete properties is the high cost of data collection. As a result, we often deal with small datasets that make training large-scale neural networks impractical.

To overcome this, we chose a different path: an ensemble of affine models. By imposing constraints on the parameters of these models, derived from the underlying physics, we created a solution tha, by construction, respects important sanity checks on the model behavior. This approach not only optimizes resources but also ensures the model’s reliability, giving us better predictions without the need for massive data.

Results

We delivered a Python library that can be called through a web endpoint. In a pilot case study, we found that when optimizing for the equivalent CO2 emissions, a reduction of 10-20% could be achieved over the best mixture in the given dataset, without any reduction in compressive strength.