At one of the world's leading oilfield services companies, headquartered in the United States, even a small mistake during well cementing can result in catastrophic financial losses. For decades, laboratory testing served as the industry's most reliable safeguard.
Thousands of laboratory tests were conducted each month. Yet caution created its own paradox: the more tests performed, the higher the costs and the longer the timelines. As operations expanded, the traditional approach gradually became a bottleneck.
Facing mounting pressure, the company's engineering teams turned to AI, leveraging more than 2.2 million historical test results accumulated over the years to build a model capable of predicting cement performance. Despite extensive efforts, however, model accuracy plateaued at around 60 per cent. That level may be acceptable in many industries, but not in oil and gas, where even a marginal deviation can trigger outsized losses.
The challenge soon reached another impasse: data volumes continued to grow, yet translating that data into real operational value remained elusive.
Solving the data bottleneck
When approaching the challenge, FPT's engineering team did not start with algorithms. They started with the data.
“After analyzing more than 45 million raw data records, we realized the core issue was not the algorithm itself, but data quality,” said Mr. Dinh Vu Quoc Trung, a representative of FPT's AI engineering team. “Previous models had been built on fragmented and inconsistent datasets, making it difficult for AI accuracy to surpass a certain threshold.”
With this insight, the team focused first on cleaning, standardizing, and restructuring the data. After three months, approximately 9 million high-quality records had been created, establishing a reliable foundation for model training.
Once the data bottleneck was resolved, model performance improved rapidly. Accuracy reached 79 per cent, meeting the partner's required threshold and enabling deployment in real-world environments. Instead of testing thousands of possible combinations in the laboratory, the AI system now pre-screens low-potential formulas, allowing engineers to focus on the most promising options.
Flipping the equation: AI begins designing solutions
The deployment of the AI model appeared to mark the end of the challenge. In reality, it was only the beginning.
Previously, engineers defined formulas while AI predicted the outcomes. Now, the process would be reversed: given operating conditions, technical specifications, and cost constraints, the system had to identify which material configurations were most worthy of testing. In other words, AI had become an active participant in the solution-design process. This proved far more complex, as the number of possible material combinations could reach tens of thousands.
Rather than evaluating possibilities sequentially through conventional methods, FPT engineers leveraged years of accumulated historical data to train the AI model. By identifying patterns often invisible to human observation, the system could rapidly narrow down configurations with the highest probability of success.
Evaluation cycles that once stretched across days or even weeks can now be reduced to minutes, a critical advantage in offshore operations, where every day of rig time may cost millions of dollars.
Beyond speed, AI is helping bring greater consistency to engineering decision-making. Instead of relying solely on individual expertise or intuition-driven assumptions, the system can simultaneously analyze vast datasets to uncover complex interactions among material components, factors that might otherwise be overlooked under operational pressure or compressed timelines.
More importantly, the AI development process itself serves as a form of knowledge preservation. Decades of engineering expertise and decision-making practices, accumulated across tens of thousands of wells, are digitized and embedded into AI models.
This carries particular significance as the oil and gas industry faces a wave of retirements among veteran specialists and a growing shortage of highly skilled talent. By embedding domain expertise into AI systems, companies can reduce the risk of institutional knowledge loss while creating a foundation for future generations of engineers to inherit and build upon.
Moving into the core of high-stakes industries
Years of working with some of the world's largest energy companies have gradually shaped how FPT approaches AI in industrial settings. What began as project-based expertise has evolved into Flezi Nergy, a suite of AI solutions built specifically for the energy sector.
“In energy, AI's value does not lie in replacing experts,” said Mr. Le Hoai Bao, FPT Software Vice President and Director of Global Energy & Utilities, FPT Corporation. “Its real strength lies in helping engineers make better decisions in environments shaped by countless variables and extremely high costs of trial and error.”
“More than 25 years of working alongside the world's leading energy companies has given FPT deep industry knowledge, enabling us not only to deploy technology, but also to develop AI models for mission-critical challenges at the heart of operations and production,” he added.
Viewed more broadly, the project also reflects a company's progression toward AI maturity through CASAN, FPT's AI-native capability framework for assessing and guiding enterprise AI transformation.
Under this model, AI is no longer treated as a standalone tool. Instead, it becomes embedded in how companies design solutions, support technical judgment, and integrate intelligence into day-to-day operations.
Google translate