Machine Learning For Steel Smelting

There are many areas of business, like heavy industry, that are seemingly far removed from automation and the introduction of AI. This is often because regular work operations in this field involve solving complex tasks which professionals, like steelmakers, are perfectly capable of doing.

However, there’s always room for improvement, so today I bring you a story of how a steel-making plant implemented AI to save money and materials when smelting steel.

Steel and AI

The service described in this story calculates the minimum required proportions of consumables (additives)— chemicals required to produce a certain type of steel.

It is crucial to get the right ratios of these chemicals in the final product, as it is the only characteristic differentiating different grades of steel from each other.

The process of smelting steel includes many variables, which all can affect the final amount of additives in steel — the amount of slag, temperature, etc. One can add a certain amount of an additive to steel, but not get that same amount of it ad the end as many chemical reactions take place during smelting, and some can affect the amount of additives left in steel after smelting.

The steelmaker is also faced with a basic principle of any business — spend the least amount of money on production and create products in the least amount of time.

The main cost of producing steel lies in the cost of additives. Decreasing the amount of additives used in smelting results in significant savings. To do this, one needs to have a sophisticated understanding of the entire steel smelting process, what factors affect the final product and how.

The steelmaker in production is engaged in predicting the content of chemical elements in the composition of the melt, relying on a huge number of factors, and makes calculations on the fly in order to comply with the restrictions and not go beyond the limits of requirements.

Chemistry for dummies

One of the chemicals steelmakers often work with is manganese. It is easy to predict its amount in the final product, and it will be used for further explanations. The system itself, however, works with many other chemicals.

Manganese almost doesn’t burn off during smelting. If you add no manganese, its amount will not increase during smelting, and vice versa — the growth of manganese in steel can happen if it is added to it.

From now on, ferro silico manganese (FeSiMn), a typical deoxidizer and alloying element, will be written about.

The increase in the manganese content with the addition of FeSiMn can be represented by the following formula.

The increase in manganese is proportional to the added FeSiMn with a certain coefficient. All that remains is to select the coefficient, after which the growth can be predicted.

The steelmaker assesses the coefficient about the same way each time. Experienced steelmakers know the coefficients for all additives they use.

The formula shows a linear increase in manganese. However, it turns out the real reaction takes place differently.

If we superimpose real data on the increase in manganese content on the graph of our formula, we will see a slightly different picture.

Take a look at the blue dots. With the same amount of added material, the increase can be different. The error in some cases is more than 0.1.

Why? The different chemical reactions taking place during smelting affect the amount of manganese, as does the unaccounted-for gain from the furnace slag.

Black dots are cases where no material was added, but there was an increase. In these cases, the steelmaker sees that a lot of furnace slag got into the melt , and logically expects that the amount of manganese will grow from the slag.

As a result, a simple formula is replaced by a more complex one, with several coefficients at once.

The system predicts what the steel composition will be, depending on the decisions made by the steelmaker. The system makes recommendations on how many materials to add, calculates the required minimum, which allows you to meet the requirements and save as much additives as possible.

Modeling in practice

The system has two parts: predictive and advisory.

• Predictive — calculations of chemical increases in certain conditions
• Advisory — selection of chemicals and their amounts based on the predictive system

There are two ways to create a predictive model:

1. Develop the most simple but understandable model with relatively low forecast accuracy.
2. Choose a complex model that operates on a large amount of data, but has other disadvantages.

Simple model

The increase in the manganese content can be estimated by a physical formula. The result is a simple, but quite an effective model, which immediately gives a very good result for some of the smelts.

The x-axis of the graph shows the real increase in Mn, and the y-axis is the predicted increase by the model. The ideal model predicts the result one-to-one with the actual gain with a straight-line graph with a 45-degree slope represented by a black dashed line.

The blue dots show smelts with FeSiMn added and the simple model predicts the increase well. The problem is black dots — cases when the increase occurred without the addition of materials, due to slag. The simple model does not cover these cases.

Complex model

Another option is to take a lot of different data and features, and then transfer it to a powerful machine learning algorithm, for example, gradient boosting. It is a popular algorithm that has proven itself to be a good tool for modeling complex nonlinear processes in many areas.

At first glance, the model looks great. It handles slag melts and is slightly more accurate in simpler cases. However, practice has shown that a complex model does not work either.

If the data on the FeSiMn addition with different values are fed into the model, the forecast will change nonlinearly, and in some places it will lose all meaning. This model is difficult to interpret.

Hybrid model

The final solution was a hybrid model consisting of two parts — simple and complex. The hybrid model receives less data, the data about the planned additions of elements is not transferred on purpose as the model needs to be able to predict the increase not explained by the addition itself.

The complex model only shifts the prediction of the linear one, and the resulting graph lacks the slope mentioned in an earlier example. This has proven to be the most efficient solution for simulating the steel-making process.

Model evaluation

Recommendation systems need to be assessed in terms of quality. The best way to do so is to conduct tests in real plant conditions, but this is not always possible. So there’s a need for analytical assessment.

Using historical data, one can ‘look into the future’ with the help of historical data. This means that it is possible to create a model that will ideally predict the increase in Mn from the addition. Such a model will not be able to work in a real setting, since it will not know the future. But in the case of historical data, this is not difficult. Real data, containing the real increase of slag, is used instead of the forecast, and as a result the forecast completely matches with reality.

This model calculates the “ideal addition” to meet the requirements. Next, it is necessary to assess what potential effect is achievable with such an “ideal model”.

The x-axis is the Mn content of the steel. The y-axis is the percentage of cases when the value on the chart was below the specified value. The gray dashed line shows how steelmakers get to the desired Mn values in a real process. In 50% of cases, for a particular steel grade, they fall about 0.51 or lower, and it is required to fall in the range from 0.46 to 0.55. The closer to the left border this line is, the more optimally the material is consumed.

The blue line shows the forecast of the ideal model, which was named Oracle. Using this model, ideal recommendations were calculated and a forecast was obtained. Oracle knows how to look into the future. In this case, the model has been adjusted so that it falls into 0.50, so the graph shows a vertical segment. A shift to the right of the vertical segment shows too much slag gain or hits the target values without recommendations. The blue line visualizes the maximum effect that can be achieved with the selected goal. You can move the target and get big savings.

The last step is to compare the chosen model with practice and with Oracle. The chosen model is the orange line on the chart. Recommendations from the model were received and passed on to Oracle to calculate what the gain would actually be. As you can see, it is close to the blue (ideal) model and gives an excellent result for a given goal.

Conclusion

Machine learning algorithms can be of great help in metallurgical production. Choosing the right model makes it easier to manage production processes, reduce the consumption of expensive ferrous materials, reduce smelting time, stabilize the chemical composition of products and increase processing productivity. Reducing the cost of ferroalloys consumption for smelting can reach 3–5%, which in absolute terms is a rather large amount.

Machine learning tools are able to perfectly solve problems in heavy industry, the main thing is to use them in suitable areas, apply a creative approach and qualitatively evaluate the final result.