CatBoost cut ferromanganese dosing error by 60% in simulated steel heats

A steelmaking supervisor adds ferromanganese to steel to achieve the required manganese content. Adding too much wastes material; adding too little risks defective steel. In an educational project on Habr, Ramil Maksyutov demonstrated how CatBoost can calculate the minimum addition using data available before the ferroalloy is added. On artificially generated heats, the model reduced mean absolute error by roughly 60% compared with a simple median prediction.

For developers and product managers, the value lies in how the solution is structured: which data the model can legitimately use, what to compare its results against, and why an attractive potential savings figure cannot immediately go into a sales proposal. The project traces the path from a spreadsheet to a calculator while also defining the limits of such a prototype.

How the model calculates the addition

The task centers on producing 35L steel in an electric arc furnace. In the simulation, the liquid steel weighs around 3000 kg, and the calculation targets a manganese content at the lower threshold of 0.45% specified by the author. This requires estimating how much of the ferromanganese the metal will actually absorb.

The same addition produces different results under different conditions. In the data generator, recovery depends on temperature, oxygen, and slag basicity. The latter describes the slag's chemical composition. The supervisor does not know the final recovery in advance, so the model receives indirect features: the masses of the charge components, oxygen consumption, temperature, basicity, and carbon content.

Imagine a calculator: a process engineer enters the heat parameters, receives a calculated minimum, then adds a safety margin. The interface can compare this dose with typical consumption. But the size of that margin needs its own justification: a slider alone does not turn a prediction into a safe production recommendation.

A crucial detail is that the features exclude the final steel composition and actual manganese recovery. These values become available after the dosing decision. Using them in training would produce a convincing test that could not be replicated when choosing a dose in practice.

What a 60% reduction in error means

The project uses 2000 synthetic heats. Of these, 1280 are used for training, 320 to select the stopping point, and another 400 for final evaluation. This split resolves a specific problem in the previous version: the test set was used both to select the number of trees and to provide the final performance estimate.

On the separate test set, CatBoost achieved a mean absolute error of 0.352 kg. A simple baseline that always predicts the training set median had an error of 0.873 kg. The difference amounts to a reduction in error of roughly 60%. This is a useful comparison: without a baseline, it is difficult to judge whether the result justifies making the system more complex.

However, the model reconstructs relationships that the author built into the generator in advance. Slag basicity accounts for 69.7% of feature importance, but that figure cannot be applied to a real furnace. In the generated data, temperature and oxygen rarely reach the ranges where their respective penalties kick in. The resulting feature importance distribution primarily describes the experimental setup.

Why almost 10 kg does not yet count as savings

In the simulation, the average actual addition is 26.9 kg, while the calculated minimum is around 17 kg. The author obtains an average difference of 9.84 kg per heat and explicitly calls it an upper estimate. The article provides no confirmed production savings.

The reason is straightforward: the target is calculated precisely at the lower permissible limit. If the model underestimates the required dose, the safety margin may be insufficient. A mean error of 0.352 kg does not reveal how dangerous individual misses are or how often they occur. Deployment will require assessing these specific cases, as well as accounting for measurement errors and changing production conditions.

The practical takeaway for a team building a similar product: start with input data availability and a simple baseline, then validate the model on real heats, taking their timing into account. The safety margin and acceptable risk must be agreed with process engineers. The article's Streamlit calculator is suitable for discussing the logic and demonstrating a prototype; a separate validation process must establish the basis for production dosing.