Q: What are the uses of concrete and what properties make it the preferred building material?
Olivetti: Concrete is the dominant building material in the world, with annual consumption of 30 billion tons. This is more than 20 times greater than the second most frequently produced material, steel, and the scale of its employ leads to a significant environmental impact, equivalent to approximately 5-8 percent of global greenhouse gas (GHG) emissions. It can be manufactured locally, has a wide range of structural applications and is cost-effective. Concrete is a mixture of fine and rough aggregate, water, cement binder (glue) and other additives.
Q: Why is it unsustainable and what research problems are you trying to solve with this project?
Olivetti: The community is working on several ways to reduce the impact of this material, including using alternative fuels to heat the cement mix, increasing energy and material efficiency, and sequestering carbon dioxide in manufacturing plants, but one essential opportunity is developing an alternative to the cement binder.
Although cement makes up 10 percent of concrete’s weight, it accounts for 80 percent of its greenhouse gas footprint. This impact is due to the burning of fuel to heat and carry out the chemical reaction required in the production process, but also the chemical reaction itself releases CO2 from limestone calcination. Therefore, partial replacement of cement inputs (traditionally ordinary Portland cement or OPC) with alternative materials from waste and by-products can reduce the greenhouse gas footprint. However, using these alternatives is not in itself more sustainable because the waste may be transported long distances, increasing fuel emissions and costs, or may require pre-treatment processes. The optimal way to employ these alternative materials will depend on the situation. However, due to the huge scale, we also need solutions that will allow the employ of huge amounts of concrete. The aim of this project is to develop pioneering concrete mixtures that will reduce the impact of cement and concrete on greenhouse gas emissions by moving away from trial-and-error processes to more predictable ones.
Chen: If we want to fight climate change and improve our environment, are there alternative ingredients or reformulations we could employ to emit less greenhouse gases? We hope to find a good answer with this machine learning project.
Q: Why is it essential to address this issue now, at this moment in history?
Olivetti: There is an urgent need to address greenhouse gas emissions as decisively as possible, and the path to doing so may not be straightforward for all industries. In the case of transport and electricity generation, paths to decarbonization of these sectors have been identified. We must act much more aggressively to achieve these goals in the time necessary; in addition, the technological approaches to achieve this goal are more see-through. However, for sectors that are hard to decarbonize, such as industrial materials production, decarbonization paths are not as well-defined.
Q: How do you plan to solve this problem to produce better concrete?
Olivetti: The aim is to predict mixtures that meet performance criteria such as strength and durability, and that balance economic and environmental impacts. The key to this is the employ of industrial waste in mixed cements and concretes. To do this, we need to understand the reactivity of the glass and mineral component materials. This reactivity not only defines the limit of possible employ in cementitious systems, but also controls the processing of concrete and the development of strength and pore structure that ultimately control the durability of concrete and the life cycle of CO2 emissions.
Chen: We are exploring the possibility of using waste materials to replace part of the cement component. We hypothesized that this would be more sustainable and economical – in fact, waste is common and costs less. Due to the reduction in cement consumption, the final concrete product would be responsible for significantly less carbon dioxide production. Determining the right proportion of concrete mix that will ensure the durability of the concrete while achieving other goals is a very hard problem. Machine learning gives us the opportunity to explore advances in predictive modeling, uncertainty quantification, and optimization to solve a problem. We explore the possibilities using deep learning and multi-objective optimization techniques to find the answer. These efforts are now more feasible and will produce the results along with the reliability estimates we need to understand what constitutes good concrete.
Q: What kind of artificial intelligence and computational techniques do you employ for this purpose?
Olivetti: We employ artificial intelligence techniques to collect data on individual concrete ingredients, mix proportions and concrete performance based on literature through natural language processing. We also add data obtained from industry and/or high-throughput atomic modeling and experimentation to optimize concrete mix design. We then employ this information to gain insight into the reactivity of possible waste materials and by-products as alternatives to cement materials to achieve low CO emissions2 Concrete. By incorporating general information about specific components, the resulting specific performance predictors are expected to be more stalwart and transformative than existing AI models.
Chen: The ultimate goal is to determine what ingredients and in what quantities should be added to a concrete production recipe that optimizes various factors: strength, cost, environmental impact, efficiency, etc. For each of the goals we develop, we need certain models: we need a model to predict the properties of concrete (e.g. how long will it last and how much weight can it bear?), a model for estimating costs and a model for estimating the amount of carbon dioxide generated. We will need to build these models using data from literature, industry, and laboratory experiments.
We study Gaussian process models to predict the strength of concrete in the following days and weeks. This model can also give us an estimate of the forecast uncertainty. Such a model requires providing parameters for which we will employ another model for calculations. At the same time, we also explore neural network models because we can inject domain knowledge from human experience into them. Some models are as uncomplicated as multi-layered perception, while others are more convoluted, such as graph neural networks. Our goal is for the model to be not only exact, but also stalwart – the input data is clamorous, and the model must account for the noise in order for its predictions to remain exact and stalwart in multi-objective optimization.
Once we have built models that we are confident in, we will apply their predictions and uncertainty estimates to multi-objective optimization, under conditions of constraints and uncertainty.
Q: How to balance the cost-benefit trade-off?
Chen: The many goals we consider are not necessarily consistent and sometimes contradict each other. The goal is to identify scenarios in which we cannot simultaneously continue to push the values of our goals without compromising one or more of them. For example, if you want to reduce costs even further, you will likely have to suffer in terms of efficiency or environmental impact. Ultimately, we will forward the results to policymakers who will analyze the results and consider available options. For example, they may be able to tolerate slightly higher costs while significantly reducing greenhouse gas emissions. Alternatively, if the cost varies slightly but the performance of the concrete changes dramatically, say doubles or triples, this is definitely a favorable outcome.
Q: What challenges do you face in this job?
Chen: The data we get from industry or literature is very clamorous; concrete measurements can vary significantly depending on where and when they are taken. When we integrate them from different sources, there is also significant data missing, so we have to put a lot of effort into organizing the data and enabling it to be used to build and train machine learning models. We also explore imputation techniques to replace missing features, as well as models that tolerate missing features in our predictive modeling and uncertainty estimation.
Q: What do you hope to achieve with this work?
Chen: Finally, we suggest one or more specific regulations or a continuum of regulations to producers and policymakers. We hope this will be invaluable information for both the construction industry and efforts to protect our beloved Earth.
Olivetti: We would like to develop a stalwart way to design cements that employ waste materials to reduce CO emissions2 footprint. No one is trying to create waste, so we can’t rely on one stream as a feedstock if we want it to be massively scalable. We need to be versatile and stalwart to adapt to changes in raw materials, and for this we need better understanding. Our approach to developing local, vigorous and versatile alternatives is to understand what makes these wastes reactive, so that we know how to optimize their employ and do so as widely as possible. We do this by developing a predictive model using software that we have developed in my group to automatically extract literature data on over 5 million texts and patents on a variety of topics. We combine this with the innovative capabilities of our IBM collaborators in designing methods to predict the final impact of up-to-date cements. If we are successful, we can reduce emissions of this ubiquitous material and play our part in achieving greenhouse gas reduction targets.