FAQ and Details

You don't find the answers you were looking for? Shoot us an email and we will take care of it.

General Questions

Do I need to complete POC1 and POC2 or can I pick one of them? VIEW

Submitting the deliverables to POC1 is mandatory – submitting the deliverables to POC2 is optional. Qualitative aspects to play a role here. For example if two teams achieve similar scores on POC1, their results on POC2 will be taken into consideration to determine the higher-ranking team.

How exactly will you evaluate my submission? VIEW

At submission deadline we will quantitatively measure your POC1 model’s performance using the test set and comparing your model’s output to the pre-existing PCA. In detail: We will use the confusion matrix coming from your model, compare it to the pre-exiting PCA’s confusion matrix.

Teams will be ranked by their performance by the jury.

In addition to the quantitative performance in POC1, the jury will apply the following qualitative criteria:

  • How well did the participant or the team document and explain their approach on POC1? The jury will evaluate the submitted documentation.
  • Did the participant or the team come up with new and non-trivial insights on POC2? If yes, how well are the insights documented and explained?

The exact key of evaluation criteria for the presentation deadline will be published shortly. Thank you for your patience.

Do I have to be a long-time data scientist to participate? VIEW

No. We are looking for creative minds with any level of expertise. Especially in our POC 2, you might easily stumble upon some great insights that we have not seen or considered yet.

Do I need expertise in the steel industry? VIEW

No. All the information you need to know will be provided. But if you are interested in any explicit further details, simply shoot us an email.

Why don’t you just hire a data scientist instead? VIEW

Because we want to explicitly be open for input from different angles, different fields and different minds. We are considering hiring data scientists though, and it is not improbable that talented participants will be invited for job interviews.

How can I participate as a team? VIEW

Please let each team member register and sign up to our challenge on this website. Then fill out the Team Registration Form to register your team. You will need to assign a Team Lead who will be responsible for submitting your team’s results on submission deadline.

Where do I find the data and all the specific information I need? VIEW

In your profile, there is a link to a Google Drive Folder. This folder contains the data, as well as a link to a Google Doc that provides all the relevant information. If you feel like you are missing some information, feel free to contact us at anytime. Any important questions that come up will also be added to this FAQ section.

Do you accept late submissions? VIEW

Be on time. There will be no exceptions for late submissions. Deadline is November 18th, 24:00 h CET. Best make sure to start your upload early in case your internet connection is slow or any unforeseen problems occur.

Is there any image data? VIEW

No. We do not provide any image data like x-ray or ultrasound.

Will there be a leaderboard? VIEW

During the challenge, there will be no leaderboard but after the submission deadline, we will let the Top 10 know how well they did on the POC 1 in comparison to the rest of the Top 10. That does not yet resemble your final rank, it only gives you an orientation where you stand. You will still be able to convince the jury qualitatively with your approach on Presentation Day – also and especially with respect to your findings regarding POC 2.

How are POC 1 and POC 2 weighted to determine the quality of my overall result? VIEW

The jury will select the Top 10 both from a quantitative perspective (how well did you outperform the PCA) as well as from a qualitative perspective. Here the POC 2 can play an important roll if you came across relevant insights we have not thought about yet (i.e. anything that is non-trivial). If these insights are significant, you might well make the Top 10 without even beating the PCA in POC 1.

What happens if I make it into the Top 10? VIEW

In a video call, you will present your approach to our jury who then qualitatively decide who wins 1st and 2nd prize. Additionally you can attend the award ceremony in Düsseldorf and afterwards go on the steel plant tour.

How are the winners selected and what is the "Presentation Day” about? VIEW

Based on your submissions, our jury will select the Top 10. These participants will be notified and will then have a video call with the jury in which they can present their approach and answer questions. After having heard all presentations, the jury will qualitatively decide who wins 1st and 2nd prize. For the preparation of the presentations we will assist you the best we can and make sure you don’t feel uncomfortable.

What if I don’t have any time on Presentation Day? VIEW

In case you don’t have any time for a video call on the 5th of December, we will have additional slots on the 6th of December.

How do you make transparent who won and why? VIEW

The recorded presentations of the 1st and 2nd place will be made available to all participants.


Why can slabs from different strands be in the same sequence? VIEW

The data comes from a two strand casting machine. One sequence contains slabs of strand 1 and 2. Both strands are independent from each other. Therefore it is pretty normal that slabs from strand 1 and strand 2 will be in the same sequence. They may even overlap with slabs from the other strand. Therefore any measurement may be totally different from the other strand. Even the ActCastLength. There is simply no connection between the two strands besides the ActLadleWeight and the ActTundishWeight, since both strands are fed from the same ladle and tundish.

When will the test set be published? VIEW

The test set is now available. It has been uploaded to the POC 1 folder in your Google Drive.

How should my submission for POC 1 look like? VIEW

Besides your Defect.csv table as the prediction result to the test set, we might need you to provide us with a virtual machine with the executable environment. If we consider it relevant, regarding your Defect.csv table, we will ask you to provide us with the VM. In case you need help setting it up, please just let us know and we will assist you.

How is the POC 1 data set structured? VIEW

The provided training data set consists of

  • 136 slabs in total
  • 11 slabs with defects
  • 174 sensor attributes

All further information will be provided in the documentation you can access with the data. But feel free to contact us if you need to know anything specific that you feel is missing.

How can I get feedback on the quality of my results? VIEW

For the PCA 1, the training set already includes whether the PCA which you are aiming to beat found the defect or not. If you have questions, we are always there for you. But we will not be able to provide you with explicit feedback on your submitted results before the submission deadline. After the submission deadline we will announce how many submissions we received in total and how many were better than the PCA.

What causes the metering roller to slip and how often? VIEW

Especially for the first slab of a sequence there are sometimes deviations caused by the start of casting processes. That is why in a few cases: SLABS.EndLength – SLABS.StartLength != SLABS.Length.

Are the descriptions of the crack positions compatible with the coordinates used for “MouldSidesCoordninates”? VIEW

Yes, that is correct.

Which parts of the cast does a measuring window display? VIEW

The data is saved time-continuously. For a measurement ActCastLength =MEASUREMENT.ActCastLength m, the lengths MEASUREMENT.ActCastLength + SENSORS.YPos are visible at the same time at the respective thermocouples.

What are the False LFC alarms about? VIEW

There are situations in the casting process like Steelgrade Change, Tundish Change or Slowdowns in which false alarms were produced because the temperatures indicated an assumedly typical longitudinal crack behaviour. In these cases, the SMS longitudinal crack detection is deactivated.


How is the POC2 data set structured? VIEW

The provided data set consists of:

  • 41K pieces from 13,5K slabs
  • 88 attributes
  • 10% inspected slabs
  • 50% defects

All further information will be provided in the documentation you can access with the data. But feel free to contact us if you need to know anything specific that you feel is missing.

How can I get feedback on the quality of my results? VIEW

For the PCA 2, there won’t be any feedback, since it is a qualitative challenge. But feel free to ask us if you think you need any type information.

You don't find the answers you were looking for? Shoot us an email and we will take care of it.