The State of Machine Learning in Concrete

In 2010, machine learning in concrete was a curiosity. A handful of papers a year, mostly theoretical.

Today it is one of the fastest growing topics in construction materials research. And the growth is not steady. It is the curve you see when a technology stops being interesting and starts being inevitable.

Why should anyone pouring concrete care about paper counts? Because research volume is a leading indicator. The methods proven in these papers reach your site three to five years later, in maturity systems, in mix optimization, in QA automation. This curve says the next five years of concrete technology are already being written. But the advantage won't go to whoever reads the papers first. It goes to whoever puts the emerging insights to work in the field first, in a landscape that is changing faster than most of its players.

Across every year of the dataset, one topic dominates: strength prediction.

Not robots. Not exotic materials. The question researchers keep pointing machine learning at is the same one a site engineer asks at 6am: is it strong enough yet?

That is validation, not coincidence. Strength timing is where the money is. Striking formwork, stressing tendons, sequencing the next pour: every one of those decisions waits on strength. And every day spent waiting for a cube result that field data could have answered is a day on the critical path.

The research community has effectively voted on where machine learning creates value in concrete. It voted for the exact problem site teams feel most.

Here is the finding that should stop you scrolling.

Research on low carbon concrete is enormous. It is the biggest topic in the field. Research on machine learning in concrete is exploding, as you have just seen. But the overlap, the papers that use machine learning to solve the carbon problem, is a sliver.

The two research worlds barely talk to each other. One asks how to predict concrete performance. The other asks how to cut concrete's emissions. They publish in different journals, cite different work, and rarely meet.

On a real project, they are one question. Every mix decision is a strength decision and a carbon decision at the same time. Pick a low clinker mix and your strength gain slows down. Chase early strength and your embodied carbon climbs. You cannot optimize one without understanding both. And the research world, for all its growth, has barely started connecting them.

The most advanced work in that intersection is not happening in journals at all. It is happening in the field, inside a small number of teams, ours among them, training models on live data from real structures. The papers will catch up. The buildings won't wait.

So where does this go next? The two streams are already bending toward each other. Fields like this don't merge gently. When prediction research and carbon research finally combine, the joined field won't grow at the pace of either parent. It will take off, the way machine learning itself did in 2019.

Every few years, researchers upgrade their favorite prediction method.

In 2016, most papers used a technique called Extreme Learning Machines. By 2020 it had all but disappeared. Fuzzy logic had a moment around 2021, then faded. Today's favorite is gradient boosting. Each of these is a different mathematical recipe for the same job: take data about a concrete mix, and predict how strong it will get. And each generation has been a genuine step forward.

The charts below show a field iterating fast, and that iteration is healthy. But watch the size of the steps. A better recipe applied to the same ingredients buys a percent of accuracy here, a percent there.

Here is why. Through all that progress, the data underneath barely changed. Most of this research still trains on a few hundred lab cube results, crushed under perfect conditions. A prediction model is only as good as what it has seen. If it has only ever seen the lab, it has never seen a cold pour in November, a truck that sat in traffic for an hour, or a supplier quietly switching cement source mid project.

The algorithms are the engine, and the engine keeps improving. But data is the fuel. Feed the same algorithms rich, real-world data at volume, thousands of instrumented pours in real weather with real supply chains, and the gains stop being incremental. That is the step change waiting in these charts, and it doesn't come from the next clever model. It comes from combining the best models with data no lab can produce.

Here is what the research world hasn't noticed yet: these aren't trade-offs. They're a cycle.

Predict strength accurately and you stop over-specifying mixes "just to be safe". Right-sized mixes use less cement, so they cost less and carry less carbon. Live field data proves quality as you pour, so there's less waiting, less rework, fewer disputes. And every instrumented pour adds to the dataset, which makes the next prediction sharper. Around it goes, and each lap makes the next one faster.

But a flywheel only spins if something turns it, and in concrete that something is operations, not data science. The prediction gains in the research literature stay theoretical until the plumbing exists on site: strength streaming from the pour in real time instead of waiting on cubes, delivery tickets and lab results landing in one clean record automatically instead of being chased by email, and forecasts arriving before the mix decision, not after it.

Walk around the wheel above: each stage feeds the next, and skipping one stops the loop. No live field data means nothing to learn from. No organized records means no usable dataset. No prediction engine means the data never changes a decision. The operational gains alone pay for the trip. The compounding is what you stay for.

Three conclusions fall out of the data. Each one changes a decision you are making right now.

This is the gap we built Mix AI to close: a predictive engine that forecasts strength gain and embodied carbon across your approved mix options before the pour. It is calibrated on live sensor data from more than a thousand real projects, not on lab datasets. We didn't wait for the two research worlds to meet. Our customers couldn't.