Episode 22: Cement clinker phases predicted with ML model

In this podcast episode, MRS Bulletin’s Laura Leay interviews Anoop Krishnan from the Indian Institute of Technology in New Delhi, India, about a machine learning model developed after a two-year period of collecting data from the cement industry, supported by the Cement and Concrete Research Network. Krishnan’s work resulted in a model that predicts the alite, belite, and ferrite content in the clinker produced by a given cement plant. These phases control cement quality and give strength to the cement over different curing times. This work was published in a recent issue of Communications Engineering.

Transcript

LAURA LEAY: Welcome to MRS Bulletin’s Materials News Podcast, providing breakthrough news & interviews with researchers on hot topics in materials research. My name is Laura Leay. Most of the time we work on samples developed in the lab, using carefully controlled conditions. Working on data taken from real industrial plants is much more challenging, but can lead to tremendous insights for materials processing and sustainability. Supported by the Cement and Concrete Research Network, data were collected over a two year period, and included cement clinker phases, process parameters such as temperatures and flow rates, and oxide compositions at various stages in the industrial process. The outcome is a model that predicts the alite, belite, and ferrite content in the clinker produced by a given cement plant. These phases control cement quality and give strength to the cement over different curing times.

ANOOP KRISHNAN: Knowing composition is one thing, but knowing the phases is a completely different thing. So unless you have the actual phases of the materials known, you wouldn’t be able to use it for anything meaningful. This can only be done by XRD – X-ray diffraction, right. But doing this is quite expensive and time-consuming and challenging. Now if you can just directly predict it so instantly you get all the phases directly from the composition, or you get all the phases directly from all the process conditions and raw materials that you’ve given, now you can optimize – basically you can play with the knobs so that you know how much C3S you want to get, how much C2S et cetera.

LAURA LEAY: That was Assistant Professor Anoop Krishnan from the Indian Institute of Technology in New Delhi, India. The model uses machine learning and incorporates expert knowledge of plant operators to gain understanding of factors such as residence time of materials in different parts of the cement plant. The datasets were inherently messy with variables recorded at different frequencies and changes made to some plants during the two-year data collection period. This meant that pre-processing was required before the data were fed into the model. 

ANOOP KRISHNAN: It’s quite interesting, because it started working for one plant but it did not start working for the other plant despite doing the same things. And that’s when we reached out to the other plant, so then they said, “Oh yes, of course, we did change our XRD machine in-between and the calibration has slightly changed”. That gave us confidence and we realized that these are at least the basic things you need to do, and from there we started moving forward.

LAURA LEAY: The model demonstrates impressive accuracy, and is shown to be superior to the Bouge calculations that are well known in the cement and concrete industry. Developed in the early 1900s, these linear equations are used to estimate the mineral phases if you know the clinker composition. In contrast, the model developed using machine learning can predict the mineral phases in the clinker using a limited number of key features of the cement plant. The model is also unique to the plant it was trained on.

ANOOP KRISHNAN: Every plant had a unique, individual problem. Even in emissions: some of them did something to control the emissions but some of them did not, some of them measured it after the intervention, some of them measured it before the intervention. So there are always very small details, so every day is like a discovery. When you work with real data, unlike a theoretical problem, nothing is under your control. 

LAURA LEAY: Member of the consortium that forms the Cement and Concrete Research Network have clearly seen the benefit of this approach. The learning from this experience can be applied to future studies and other aspects of the work can contribute to the ambition of the cement and concrete industry to achieve net zero greenhouse gas emissions by 2050. The industry contributes around 10% of global greenhouse gas emissions; not only is materials production energy intensive but carbon dioxide is produced during the calcination of limestone. While the industry transitions to green production methods, existing plants could work much more efficiently using this new modelling method.  

ANOOP KRISHNAN: This is only the first work in the series of works that we are doing. Now that we know how to make it work we have several more in the pipeline: one on emissions, one on energy optimization, and one on – for example – what are the good practices when you are collecting data or what are the good practices you should be doing when developing the model et cetera. So we have a lot of learning from this work that we are trying to document.

LAURA LEAY: This work was published in a recent issue of Communications Engineering. My name is Laura Leay from the Materials Research Society. For more news, log onto the MRS Bulletin website at mrsbulletin.org and follow us on X, @MRSBulletin. Don’t miss the next episode of MRS Bulletin Materials News – subscribe now. Thank you for listening.