Titanium Furnace Temperature & Pressure Forecasting
Time series forecasting models to predict threshold breaches in titanium smelting furnaces, enabling operators to intervene before unplanned shutdowns. Built at Rio Tinto's PACE analytics team.
Overview
I was a data scientist on Rio Tinto's PACE analytics team, the company's internal consulting group for data and ML. PACE operates like an in-house consultancy, rotating across business units to solve problems with analytics and machine learning.
The project: forecast temperature and pressure time series inside titanium smelting furnaces. When furnace readings cross certain thresholds, operations have to stop. Downtime is expensive. One internal study estimated the solution could save tens of millions of dollars a year by predicting threshold breaches before they happen and giving operators time to intervene.
Estimated Impact
Tens of millions of dollars per year in reduced unplanned furnace downtime, based on the client team's own cost analysis.
I left Rio Tinto before the solution was fully deployed into production.
What I Did
Subject Matter Expert Interviews
Interviewed furnace operators and process engineers to understand the physical process, what the sensors measure, what thresholds matter, and what operational decisions depend on the forecasts. This domain understanding shaped every downstream decision, from feature engineering to model evaluation criteria.
Data Catalogue & ETL
Identified where the relevant data lived across Rio Tinto's systems: sensor readings, operational logs, maintenance records. Mapped out what was available, what was missing, and what needed cleaning.
Built data pipelines to extract data from legacy source systems and load it securely into a cloud development environment on AWS. Security mattered. Rio Tinto's operational data is sensitive.
Feature Engineering
Engineered features from raw sensor time series and operational metadata to feed the forecasting models. This included lag features, rolling statistics, and domain-specific transformations based on what the SMEs described about furnace behavior.
Modelling
Built time series forecasting models using classical ML techniques, primarily gradient boosting. The models predicted temperature and pressure readings across different zones of the furnace based on historical patterns.
Client Presentation
Presented the prototype results and business case to stakeholders. The estimated impact came from the client team's own analysis of how much unplanned shutdowns cost.
The most important work happened before any model was trained: understanding the physical process, locating the right data, and building trust with the people who would use the forecasts.
Results
Built a working prototype that forecasted furnace temperature and pressure from historical sensor data. The internal study estimated tens of millions in annual savings from reduced downtime. I left Rio Tinto before the model was deployed to production.