Internship
Data Scientist Internship
Umicore · Olen, Belgium · Sep 2024 to Dec 2024
Between September and December 2024 I joined Umicore, a global materials technology and recycling company, in Olen, Belgium, as a Data Scientist Intern. The internship centred on one substantial machine learning project: building an automated system to classify patents related to battery technologies. The three documents at the bottom of this page, my research and planning, realization and experiments, and personal reflection, tell the full story. This page is the short version.
The Project
AI Patent Classification System
An automated system that classifies battery technology patents, moving from a simple Bag of Words baseline to fine tuned, transformer based language models.
The Problem
Umicore's research and IP teams deal with thousands of battery technology patents. Reading and categorising them by hand is slow, and the existing approach, a Bag of Words model, treats each document as a loose collection of words. It has no sense of meaning, so it struggles with the dense, technical, and multilingual language patents are written in. The goal was a classifier that genuinely understands the text.
The Approach
I started by analysing why the Bag of Words baseline fell short, which framed the case for a semantic approach. From there I implemented and fine tuned transformer based models, PatentBERT and BatteryBERT, on Umicore's domain specific data. Patents are unevenly distributed across categories, so I tackled class imbalance with data augmentation, and processed large multilingual text corpora into clean training data. I experimented with meta model embeddings to squeeze out more performance, and tracked every experiment systematically with MLflow inside Azure Databricks, so results stayed reproducible and comparable. Along the way I hardened the preprocessing scripts to make the whole pipeline more stable.
The Outcome
The fine tuned transformer model reached 81% classification accuracy, a clear step up from the Bag of Words baseline, on a pipeline that is stable, reproducible, and ready for production integration. Just as importantly, the project gave Umicore a documented, evidence backed comparison of approaches to build on.
My Role
What I worked on
- Prepared, cleaned, and validated large multilingual datasets for the NLP pipelines.
- Implemented and fine tuned transformer based models (PatentBERT, BatteryBERT) on domain specific data, raising classification accuracy to 81%.
- Tackled class imbalance through data augmentation and experimented with meta model embeddings to push performance further.
- Used MLflow in Azure Databricks to track experiments, diagnose pipeline issues, and compare model performance.
- Enhanced the preprocessing scripts to improve pipeline stability and reduce downstream errors.
- Collaborated with engineers and researchers to turn analytical findings into actionable improvements.
Tech Stack
Documents
Read the full story
The three documents below cover the internship in depth. Each one is readable right here on the page, or you can open it full screen.
Research and Planning
The plan I set out at the start of the internship, covering scope, goals, milestones, and intended approach.
Realization and Experiments
The full write up of the experiments, methods, results, and findings of the patent classification project.
Personal Reflection
A personal essay reflecting on what I learned and how I grew over the course of the internship.