Decarbonizing Concrete with Artificial Intelligence
Abstract:
With an annual production of 4 tons per capita, concrete is the second most used material in the world after water. Although concrete has largely defined modern society, it comes with a hidden cost: it is a climate killer. Concrete contributes to 8% of global CO 2 emissions, which is quadruple the emissions of the entire aviation industry. In this presentation, I will discuss how artificial intelligence can be used to reduce the carbon footprint of concrete. Based on a dataset of more than 1 million concrete mixtures, we trained a series of machine learning models that accurately predict the performance of a concrete formulation based on its mixture proportions.
Based on these models, we introduced an inverse design engine that generates optimal concrete formulations featuring minimum carbon footprint while meeting all required performance targets and constraints. This approach results in an average reduction in concrete’s global warming potential (GWP) of 30%—with no changes in the raw materials, no modification of the production process, and no cost premium.
Bio:
Mathieu Bauchy is an Associate Professor in the Civil & Environmental Engineering Department at the University of California, Los Angeles (UCLA). His research focuses on decoding the physics governing the behavior of construction materials using simulations and artificial intelligence—with a focus on decarbonizing the construction industry. He is also co-founder of the cleantech startup Concrete.ai, which uses generative AI to prescribe new concrete formulations that are both less carbon-intensive and more economical.
Abstract:
Focus: decarbonizing concrete
Concrete:
Critical material for most construction: buildings, roads
4 Tons per Capital Annual Production: most used material after water
Significant CO2 emissions (driven by global scale, rather than per-Ton carbon intensity)
Concrete = water+sand+cement+gravel+additives
Concrete Carbon Footprint
8% of global CO2 emissions (e.g. aviation is 2%)
Hard to decarbinize
Stages:
Production
Construction
Use
End of Life
Focus: production stage / cradle-to-gate
Extraction and upstream production,
Transportation to factory,
Manufacturing
Most challenging right now; more progress has been made on later stages
50%+ of overall emissions: https://www.istructe.org/IStructE/media/Public/TSE-Archive/2020/A-brief-guide-to-calculating-embodied-carbon.pdf
Goal: attribute emissions to different stages of production
Cement in concrete
15% by mass
50% by material cost
95% by carbon footprint
Opportunity: reduce use of cement in concrete without compromising quality / structural integrity
Challenge: solution must work at the huge global scale (e.g. alternative materials must be available in a large volume)
Approach: use AI to optimize concrete design
Balance:
Product attributes: strength, durability, ec.
Manufacturing attributes: slump (how liquid it is), setting time at construction site, pumpability, etc. (affect cost of production and use of concrete)
Optimization
Cost function
Degrees of freedom
Constraints
Primary degrees of freedom: ratios of the various inputs
Fine aggregates
Coarse aggregates
Water
Air
Cement
May be replaced by alternative materials: fly ash, slag, silica fume
.8 Ton of CO2 released per ton, capturing this CO2 will increase cost of concrete by 2x/3x
Admixtures
Input availability varies by location
Challenge of optimization
1.34E10 possible solutions in design space
800 average number of concrete formulations per plant for different use-cases
No good way to simulate concrete due to its internal heterogeneity, multi-scale dynamics
Optimization via machine learning
Datasets:
Very little ~1K data points
High variability in measurement, many outliers:
(measurement error, human error, error in concrete batchng/machinery, data processing/recording error, intrinsic variability of concrete, unexpected but true behavior)
Need to extrapolate from datasets to different concrete designs that don’t exist yet
Need to explicitly capture uncertainty (e.g. in strength estimates)
Concrete database:
>1m data points
Can understand tradeoffs between water/cement ratio and strength, cement solid fraction and strength
Cement database
>2k datapoints
Different chemical compositions
Fly ash database
>20k datapoints
Different chemical compositions
Managing outliers via an ensemble-based detector
Extrapolation from available data vie leave-one-cluster-out cross validation
Cluster data
Train on all-but-1 cluster, predict on remaining cluster
Adjust hyperparameters to minimize validation error to avoid in-sample overfitting
Ensembled neural networks to predict compressive strength, shrinkage, slump
Human vs AI competition
Goal: densign concrete mix with lowest embodied carbon showing 5000psi strength
AI succeed in
Meeting strength target and
Meeting slump goal
Lower embodied CO2
Material cost decreased
Can estimate uncertainty in its predictions
Has less bias than human, can use material combinations that human expert did not anticipate
Human barely missed on strength target, design had more CO2
SHAP analysis to understand reasons for AI’s design
From laboratory to field
Challenge: regionality of materials
Material variations over time
Large number of mix designs per plant for different use-cases
Manufacturing uncertainties
Hard to quantify properties with no standard test (finishability, pumpability)
Data availability is limited
Concrete.ai: software solution to use generative AI to reduce concrete’s cost and carbon footprint
Partners across North America
Native software integration to directly connect to manufacturer software to get data
High-throughput optimization of all concretes at a given plant at once
Active learning that dynamically adapts models
2 million cubic yards of concrete optimized
30% carbon footprint reduction (avg)
5% materials cost savings
Estimate: if everyone used AI to optimize concrete this would mitigate 500m Tons of CO2 emissions