[Side Project] Raw Material Input Suggestion for the VOD Process in Stainless Steelmaking
In the Side Project Series, I explore topics outside of work in which I apply my supply chain knowledge and/or technological skills to solve problems in new fields, and occasionally discover some managerial wisdom that can be leveraged across domains.
I am currently leading a team of three to design an AI solution for the largest stainless steel manufacturer (hereafter referred to as “YL” or “the client”) in Taiwan and the second largest in Asia. This project is a rare opportunity for me to polish my data analysis skills in an industry that is completely new to me. While the project is ongoing, I am recording my exploratory and brainstorming process for the best analysis approach here:
Stainless Steelmaking
Stainless steel is a centuries-old craft that has grown into a USD 119.3 billion industry. This capacity-driven and capital-intensive industry is dominated by a few big players. As of 2021, China holds over half of the global market share.
The manufacturing process involves a series of steps:
- Raw Material: Stainless steel is composed of iron (Fe), chromium (Cr), silicon (Si), nickel (Ni), carbon (C), nitrogen (N), and so on. The ratio of the raw material amounts determines the characteristic of the alloy and its designated “grade”.
- EAF (Electric Arc Furnace): The EAF is used to melt solid iron scraps into liquid form at over 1800 degrees Celsius.
- LHF (Ladle Heating Furnace): The LHF heats the steel to allow adjustments to the molten metal’s chemical composition.
- MRP (Metal Refining Process): The MRP carries out decarburization using oxygen and inert gases.
- VOD (Vacuum Oxygen Decarburization): The VOD process is used to lower the carbon content without compromising the Cr yield, by inducing argon from the bottom and oxygen from the top. The alternative method is the AOD method, which uses argon rather than pressure adjustment to refine the molten steel.
- CC (Continuous Casting): CC is the process of infusing melted steel in a certain mold form.
We can break down the cost of manufacturing into roughly two parts: raw materials and fabrication. The former accounts for the majority of the production costs (around 70–80% according to this research, excluding the procedure to get the steel to its final form). Of course, the post-pandemic raw material cost spike only underscores the urgency to make the most efficient use of every resource.
The Problem
Based on the above, it is evident why the client needs our help to search for the best material combination to input, with a focus on the VOD process, to reduce current production costs.
Studying the Data
We have been given production records for the past four years to do the analysis. Each record consists of 6,000–7,000 rows of data. Data includes:
- Production profile: date, shift, steel grade, furnace number…
- Chemical composition (C, Mn, Si, Cr, Ni, Mo, Cu, Al…etc.): initial C %, stage(n) C %, final C %, initial Mn %, stage(n) Mn %…
- Additional components: kgs of (x) alloy added at stage(n), kgs of (y) alloy added at stage(n), total amount of argon added, total amount of nitrogen added…
- Mechanical properties: mixing time, initial temperature, stage (n) temperature, oxygen blowing rate, vacuum degree, slag basicity…
From studying the datasets, we can observe some follow-up actions:
- There is a significant gap between column values. For instance, the “stage(n) C %” usually falls around 0.12–0.08, and the “stage(n) Cr %” usually falls fall around 10–15. Hence, the data needs to be normalized before being used for training.
- We need to clarify with YL’s engineers whether there are duplicate columns among the columns and whether all data is accurately recorded. Doing some exploratory analysis will also help to spot problems.
- There may be a demand for more detailed information down the road. For instance, we may need information on the amount of argon induced at each time interval, in addition to the “total amount of argon added”. Whether we can get additional data depends on the client’s willingness to share it, as well as the availability of such information.
Crafting the Best Approach
After taking a comprehensive view of the datasets, next up is selecting the best ANN model for analysis. I also considered multiple aspects of the problem to help me make a decision:
1. What solution will generate the most business value?
As mentioned before, steels are categorized into various “steel grades.” Some grades, such as the SAE 304 widely used in household and industrial applications, produce more sales than others. YL should have its top-selling products contribute the majority of their revenue or profit. Thus, it is reasonable to use that steel grade as the entry point for composition engineering.
Now, back up a little. Couldn’t we just build a model that incorporates data from all product types?
The answer is “Yes,” but the tradeoff would be a less precise resulting model because meshing all the data balances out the unique reactions of any specific grade. In an industry so mature and rigid, creating an overly generalized model probably won’t add much value to the client.
Therefore, creating multiple models, each targeted toward a specific steel grade, is the better way to go. The next step is to communicate with the client to determine the product type that is most critical to the business or holds the most potential for improvement, and explain to the stakeholder why we think the “less efficient” method is more appropriate for this case.
§ Key stakeholder: Project owner
2. What constraints (or opportunities) does the domain knowledge imply?
The VOD process contains three stages: oxygen blowing, degassing, and reducing. Each stage has its respective input and output chemical compositions, and the output composition of one stage becomes the input composition of the next.
Now, it may seem intuitive to create three models that run consecutively. But is that the right approach?
The answer is “No.” Building three models implies that there is no correlation between the stages and that only the chemical composition is passed down. However, from consulting with the domain experts, we found that the mechanical properties have a lasting impact even if we have crossed over to the next stage.
This key insight that will make all the difference in choosing the model design would have been too profound for laymen to realize without the help of experts. This example shows the importance of asking the right questions to those who understand the processes deeply and shine a light on their perspectives.
§ Key stakeholder: Engineers or domain experts
3. How will the analysis be applied to the client’s operations?
If we look at the problem proposed by the client, we might straightforwardly come up with a model that finds the relationship between input parameters and the final chemical composition. And if we go on to question, “What insights does such a model offer?” we will find that a model with the final composition as the output parameters cannot directly tell us what input parameters should be used to achieve the minimum cost.
Instead, we can take the costs and/or quality associated with each component or process into account to create a function that calculates the total cost or the health index of the production, then use the backpropagation feature of neural network (or use methods besides ANN methods, such as the genetic algorithm) to search for an optimal. Then, the model output can be applied by the client out of the box without the need for additional interpretation.
Today, AI or machine learning is seen by many businesses as a magic trick to solve all kinds of problems. Such a misconception can often lead to ambiguous problem definitions. So, it is vital to guide the client, rather than become misguided by them, to find the true pain points in their organizations and tailor a solution accordingly.
§ Key stakeholder: Project manager
My Learnings
When solving problems with AI/ML, focus on the project ROI rather than the technology itself. Ask smart questions that can usher in a solution that creates value for the client along as many dimensions as possible. Some recommended checkpoints and considerations for your AI project can be found here:
https://www.pwc.com/us/en/tech-effect/ai-analytics/artificial-intelligence-roi.html
Please leave a comment below and let me know your thoughts. If you would like to see more articles like this, please give me 10 claps!
