Rio Tinto documents 30-year-old manufacturing system using AI

Rio Tinto has built an AI domain assistant that documents the knowledge, dependencies and decision logic used by a key manufacturing system in its aluminium operations.

Rio Tinto's Ke Shi.

Data science senior advisor Ke Shi told the AWS Summit Sydney that the assistant is unlocking “accumulated knowledge and complexity” embedded in one of its most critical systems, a 30-year-old manufacturing execution system (MES) called Metpro.

Metpro, he said, manages the aluminium product lifecycle end-to-end, “from tapping right through to shipments, and acts as a central hub connecting the process control system, data capture platform and operational infrastructure.”

Shi said that over time, Metpro’s technical documentation had become fragmented, spread across “thousands” of documents. 

Additionally, the system’s highly coupled nature meant that dependencies were not always documented or well understood.

“That means even a small change like a UI configuration can have very unexpected downstream impact,” Shi said.

“[Altogether] for engineers, this translates into slow onboarding, difficulty in finding the right expertise, and high risk when making changes.”

Rather than try to rewrite or replace Metpro, given how embedded it is in daily functioning of Rio Tinto Aluminium Pacific Operations, Shi said the company instead looked for ways to “retain and operationalise decades of embedded knowledge.”

“We needed a way to preserve how the system actually works, including dependencies and the decision logic that were incomplete or had become outdated over time, and make that accessible to engineers without changing the underlying platform,” Shi said.

“Our goal was to bring together the fragmented technical documentation and turn them into something more coherent, rich and usable.”

The miner started by building a “domain-aligned training dataset”, plugging the Metpro code base and “business and operational context” into Amazon Bedrock Knowledge Bases and Amazon Bedrock AgentCore.

The company then used Amazon SageMaker AI Jumpstart - a service used to access, evaluate and compare foundation models - and adopted Llama 3.1 8B as the initial inference model.

“Using our curated dataset and Amazon SageMaker AI Jumpstart, we trained the model to internalise how the system actually behaves and more interestingly, how we expect it to respond,” Shi said.

“Fine-tuning gave the model a strong domain understanding, but the real world doesn’t stand still. Knowledge changes, processes evolve, and new information appears every single day. 

“Agents help bridge that gap by working alongside the trusted knowledge, pulling up to date information as needed, without having to retrain the model every single time.

“The agent layer keeps the understanding [of the Metpro system] refreshed, reliable and connected to reality.”

Shi said that with the domain assistant in place, engineers could understand dependencies that may impact proposed changes within the system “in minutes rather than days”.

Engineers also spent “less time reverse-engineering complex logic, [which] means more time [can be] focused on innovation and improvements.”

“Most importantly, this established our future-ready foundation. Knowledge is now captured, understood and applied in a way that supports incremental, low-risk modernisation without disrupting the critical system,” he said.

Ry Crozier travelled to AWS Summit Sydney as a guest of AWS.