Carbon Re Develops Cutting Edge Decarbonization Technology with Weights & Biases
"W&B has made it easy for knowledge sharing. I can quickly show someone an experiment I ran, the outcomes, and explain my work."
Theo Wolf
Machine Learning Engineer
Building the Modern World
Today, second only to water, concrete is the most widely-used substance in the world. The material is the foundation of modern infrastructure, mainly thanks to its durability, strength, and versatility.
Cement is the key ingredient that makes concrete such a useful building material, but it comes with a high environmental cost. About 8% of the planet’s human-produced CO2 emissions comes from cement production. To put it into perspective, if the cement industry were a country, it would be the third-largest emitter of CO2 in the world, after the U.S. and China.
One startup seeks to change that. Spun out from Cambridge University and UCL, Carbon Re aims to tackle the gigatonnes of emissions from cement production. By leveraging the immense potential of artificial intelligence, the company aims to reduce carbon emissions and costs at the same time—a win-win for all.
Cement Production
While cement is a popular and well-understood commodity product, cement production is a complex process involving several steps with ever-changing inputs (fuels, raw materials), conditions (state of equipment, shift changes), and competing priorities (throughput, regulatory requirements)
Typically, over half of the cement CO2 emissions come from the production of clinker, a major component of cement, in which limestone (CaCO3) is converted to lime (CaO) by heating to an extreme temperature. A further 40% of CO2 emissions are the result of burning fossil fuels to heat the preheater and kiln. The last 10% of emissions comes from electricity in grinding materials and transporting the raw materials.
Delta Zero Cement
Walking into a control room at a cement plant and you would be struck by the multitude of monitors surrounding the control room operators. Every decision they make directly impacts the safety, reliability, and sustainability of the facility.
Carbon Re’s first product, Delta Zero Cement, helps operators optimize how they run their kiln and preheater. The resulting increased efficiency means that the kilns use less fuel, and in turn, reduces the amount of CO2 released into the atmosphere. “What we’re doing now is empowering operators so that they can do their job more efficiently with the least amount of carbon footprint,” said Nantas Nardelli, Senior Research Scientist at Carbon Re.
The models employed in Delta Zero Cement are trained on an abundance of data derived from sensors found all over the plant. While having a large dataset is generally beneficial for training models, the fact that the team is dealing with time series data—makes matters more complicated. The data, which consists of sequences of observations recorded at regular time intervals, may contain noise, be highly lumpy, or even be intermittent depending on the context.
In the past, this continuous data stream was used for operators to make reactive decisions based on the current state of the plant. Using that data for predictive, forward-planning was not top of mind. Carbon Re is one of the first groups to maximize the value and relevance of the data, building models that are bespoke digital twins of a cement plant. This digital twin provides a platform for AI agents to learn the complexities of cement production in a sophisticated simulation. The outputs are insights on how to maximize the efficiency of the plant, enabling significant cost savings and above all—emission reductions.
Institutionalizing Knowledge and Insights
Tackling climate change requires collective action. As Carbon Re continues to recruit more like-minded individuals to join the cause, the team needs a reliable, organized, and accessible system of ML records to help onboard new members faster. With a platform like Weights & Biases, there’s a traceable history of data, experiments, models, and deployments making it simpler to reference previous projects, sometimes months or more later.
“W&B has made it easy for knowledge sharing. I can quickly show someone an experiment I ran, the outcomes, and explain my work,” said Theo Wolf, Machine Learning Engineer at Carbon Re.
Finding a way to accelerate their research is another priority for Carbon Re. Using W&B Sweeps, the team can automatically try hundreds of combinations of hyperparameter values, find the best ones, and gain a rich set of visualizations to inform their training process. This removes a lot of the manual work, freeing up each member’s time to focus on higher-value tasks.
Moving from experimentation to production, the team previously relied on manual deployments that were time-consuming and at times, resulted in incidents. That’s a thing of the past since using W&B Models. Now the team can automatically retrain and re-evaluate models to ensure the best ones are headed to production. “We use the model registry as our source of truth,” explained Nantas. “It’s a key component of our backend.”
Chasing Carbon Zero
While initiatives such as Industry 4.0 have led to the rise of data, the cement industry as a whole is still trailing behind in terms of digitalization. Legacy systems and operational processes aren’t able to make use of the available data, which is where Carbon Re comes in.
Pushing the boundaries of AI to accelerate decarbonization is by no means an easy feat. Leveraging Weights & Biases means Carbon Re can centralize and reuse knowledge, automate time-consuming MLOps tasks, and streamline the iterative modeling process. Carbon Re is committed to making a massive impact on global emissions, and Weights & Biases is thrilled to play a small part in their endeavor.
