Soft Sensors for quality prediction at Samarco

TL;DR;

Using the same template created for virtual belt scales, Samarco built a silica prediction model for flotation. With SIENTIA™’s automated retraining, even a simple regression delivers reliable, real-time insights without costly custom development.

The case

In iron ore flotation, silica content is a critical quality parameter. Traditionally, plants rely on laboratory analyses every two hours, leaving Advanced Process Controllers (APCs) operating with delayed data. This slows decision-making, increases reagent waste, and risks product quality.

Samarco addressed this challenge with a virtual quality sensor, designed to predict silica grade every 5 minutes.

Surprisingly, the model wasn’t built on complex deep learning, but on a multivariate linear regression.

 Reusing What Already Worked
 

Instead of reinventing the wheel, Samarco reused the same no-code template originally created for conveyor belt mass flow models. This was possible thanks to the flexibility of SIENTIA™, which allowed the Process Optimization team to adapt an existing template to a new use case.

But flotation is a nonlinear process, and a simple regression alone wouldn’t be robust enough over time. This is where another key feature of SIENTIA™ came into play.

Automated Retraining: The Game-Changer
 

SIENTIA™ provided automated retraining every 4 hours, which compensated for process variability and data drift. This capability transformed what could have been an unstable model into a reliable, production-ready sensor.

The combination of a simple template + automated retraining meant that Samarco could:

• Avoid exorbitant costs of custom model development.
• Leverage its own team’s process knowledge.
• Deploy a working solution much faster than traditional R&D cycles.
   

Impact on Operations

• Silica predictions delivered every 5 minutes instead of waiting 2 hours.
• Direct integration with the flotation APC, enabling real-time reagent dosing.
• Significant reduction in reagent waste and higher product quality stability.
  

This case is a good example that even “simple” models, when combined with robust MLOps features, can solve complex nonlinear problems.