Panyam Cements from Andhra Pradesh has implemented a non-linear model to formulate a better clinker composition. Sreekanth Sajjala recounts the results obtained with non-linear modelling.
Designing a better or a different clinker is a time-consuming effort. Right from the raw mix, the coal and the operating parameters of the kiln have to be changed. Often, this can impact the sta-ble running of the kiln which becomes very expensive in terms of fuel consumption and sub-optimal production.
Assessing a new clinker is also not instant, it takes 28 days for all the strengths to be measured. This means that all in all, it takes about two-three months to design and evaluate a new clinker composition. There currently are no accurate prediction methods as the formation of clinker and the link between its material properties and chemical composition are not linear. Machine learning is a field of study that gives computers the ability to learn without being explicitly programmed. Computers are not programmed but are expected to identify patterns and thereby build a model which can generate repeatable results.
These patterns need not be linear or direct, which means that machine learning methods are ex-tremely well suited in the prediction of cement properties.
Our clinker had excellent 28-day strengths (680-700) and relatively lower 1-day strengths (140-160); this meant that our customers were disappointed by our PSC where the addition of GGBS (a material with a low heat of hydration) further lowered the 1-day strengths. This meant that we couldn´t mix more than 45 per cent GGBS in PSC though up to 60 per cent was allowed. We needed a clinker which would have higher initial strengths and quickly, as the opportunity cost was quite high. We, in the R&D division, turned to non-linear modeling to design the chemical composition of a clinker that would have up to 50 per cent higher initial strengths.
The most important thing in non-linear modeling was the availability of reliable data and a lot of it. Though our lab maintained immaculate records of both chemical and physical aspects of clinker, they were all in physical form. We digitised over 18 months´ worth of clinker data. Over 500 records were used as the training set. The surface area and percentage of gypsum were not taken into account, as all the lab samples were ground evenly and had the same amount of gyp-sum.
A neural network with four hidden layers was used with a hyperbolic tangent activation func-tion. Neural networks are computer systems modeled on the human brain and nervous system. They are used mainly for non-linear modeling and computer vision. Each node is assigned a weight with which the incoming input matrix is multiplied by. These weights are assigned during training where back propagation and other training methods are used with the training set. The model is then validated on a smaller separate dataset called the testing set.
When the directional output was plotted with the various variables, the optimal composition was realised for higher initial strengths. We were able to formulate a viable clinker with over 50 per cent higher 1-day strength and 10 per cent lower 28-day strength. The next step was to design a model which could predict the kiln´s output chemical composition on the basis of the raw meal chemical composition, coal consumption and the operating parameters. We believe this could help us match the raw mix and coal more optimally and help us preserve coal and high grade limestone. One thing non-linear modeling can do which linear modeling cannot do, is factor in the indirect impact of trace compounds like MgO. Non-linear models, by observing correlations, can link trace elects to the indirect effects like their function in the formation of major com-pounds. This model has helped us increase how much GGBS we add in our PSC and has contributed to great cost savings through clinker conservation.
About the author
Sreekanth Sajjala is a part of the R&D team at Panyam Cements, Andhra Pradesh. He is one of the main contributors to the model Panyam has developed to predict the strength of clinker on the basis of its chemical composition.