Tackling a Viridien Challenge in HackTheBurgh 2026!

Neel Haval — Sun, 08 Nov 2026 11:20:35 +0000

For HackTheBurgh 2026 we tackled a track by Viridien. In this challenge, we converted unsorted lithological data consisting of charts depicting the presence of certain types of rocks at specific depths into structured outputs which may be used in data processing in industry.

Specifically, the rock columns in the charts had to be broken down into strips to identify features like intensity, colour, number of components and average distance of components which is engineered to differentiate the patterns we have. We further processed these feature vectors first to identify boundary lines (segmenting the rock column by various types). Following this, we obtain the feature vector of the segments as well as the legend images which then allows us to pair a rock type in the column with a certain classification of rocks. Our initial approach to cluster similar rock types in the rock column together failed because the accuracy of rock classification was not sufficiently high.

For testing, we started with a manual labelling to form out test set, however this was very tedious. Following this, we decided to use AI tool to analyse the images. Then we made another custom tool to label manually at least 5 examples per pattern per legend in each data set example. We then took this data set and tried to convert it to a matching pattern to obtain synthetic data set to run our evaluation on, as well as integrating publicly available data sets to make the evaluation data set more representative of the actual distribution. We made a custom CI/CD pipeline to make this evaluation process easy for everyone. Our team then decided to test the current strategy onto the evaluation set allowing us to understand where the process was failing and try different approaches to improve labelling classification.

A web application which formed the front - end was made to allow seamless analysis of rock characteristics at certain depths. Experts and researchers can analyse wells using these lithology columns and gain information such as ease of drilling, costs and capacities of various rocks to store compounds. For example if a developer wants to analyse a certain region, then they can select this area and view key characteristics such as those mentioned above regarding rock composition.

The reason those track was chosen was because it was more aligned to real applications of machine learning and computer science and was at the forefront and intersection of science, technology and industry. It gave us valuable industry experience with developing a specific use case for a client. The environmental nature of this project made it one who's impact could be measured and realised tangibly. This made our work more rewarding as we were able to see the impact that our work had. Designing methods to evaluate the solution was challenging and creating the rock classification algorithm for our strategy was confusing at times at it required constant redirection adapting.

DEV Community: Neel Haval

Tackling a Viridien Challenge in HackTheBurgh 2026!