Artificial intelligence (AI), once the stuff of science fiction, makes it possible to augment human capabilities and reimagine the future of energy.
While not all AI products live up to their hype, having a solid plan for moving from initial proof of concept to scale could help companies realize value from their AI initiatives, speakers said during the "Transforming Energy Business With AI: Vision to Reality" special session on 21 October at SPE’s Annual Technical Conference and Exhibition in Houston.
Sid Misra, associate professor at Texas A&M University, said, “Artificial intelligence may not be as intelligent as what you want it to be, but the progress that it has shown is very promising.”
Vural Sander-Suicmez, AI champion at Abu Dhabi National Oil Company, noted that AI has been present in the industry for a long time and said it is “here to stay.” He said companies can gain significant benefit from embedding AI solutions in their workflows.
Session moderator Amr El Bakry, senior principal in data science at ExxonMobil, said AI solutions don’t always live up to the hype and claims about them, but that companies using a disciplined strategy can still find success with implementations, particularly when it comes to scaling from proof of concept.
“Realizing AI’s full potential requires more than just deploying algorithms. I think all of us have been there before where you deploy a prototype concept and it just stays as such,” he said.
A three-pronged approach to scaling includes a disciplined enterprise-scale strategy that ensures value creation and continuous learning, he said. The second piece is space for innovation, which could mean reimagined businesses or workflows, he said. The third piece is the evolving AI/human interface, which includes the potential for collaboration, or a partnership between the machine and the human, he said.
“Focus on value-based deployment. To move from a proof of concept to a production-level deployment is not an easy task. It requires processes. It requires people who know how to do it. It requires multiple skills and building teams that can deliver that at enterprise scale,” he said.
Sander-Suicmez said people, governance, culture, and change management can be hurdles to scaling AI solutions. Even so, these hurdles are being overcome, he said, because the tools offer such value.
“Sometimes it’s not easy to measure the value,” he said. “Value can be time reduction. Value can be increasing the capital efficiency. Value can be increasing the health and safety performances. Value can be increasing the productivity.”
Misra pointed out another potential hurdle to scaling AI solutions.
“I think that the reason AI models are very difficult to scale, especially generative AI solutions, is they’re probabilistic. They’re based on probability, and, in critical industries like we are in, we need deterministic solutions. So that’s the challenge. How can you convert a probabilistic prediction, which is very sensitive to the prompts, to something that can be applied to deterministic critical industry requirements?” he said.
Accountability if something is wrong with a prediction is another issue, he said.
“Trying to figure out the model’s reasoning is very complex,” he added.
GIGO and Bias
Yingwei Yu, applied science manager at AWS, said the availability and quality of data for AI projects can be problematic. And a lack of high-quality data going into a model means garbage in, garbage out (GIGO), he said.
Sander-Suicmez said the industry has the luxury of years of production data from wells, and alongside that is the opportunity to convert that data into smart data.
El Bakry said company mindsets need to shift to viewing data as an actual asset. Data as an asset requires maintenance, in much the same way physical assets such as compressors or pipelines do.
“We know how to maintain them, we know what integrity means” for physical assets, he said. “If we think of data as an asset that needs continuous maintenance—not only the one-time ‘Let’s put a lot of money in it one time and it’s done’—I think that will change the way not only our industry but every industry will look at data, and that will help with the future.”
Bias is another issue that could affect the performance of generative AI, Misra said. Bias can creep in through three main avenues: bias in the data the generative AI’s model used, bias within the development team and the team’s workflow, and bias in prompts entered into generative AI tools.
“Generative AI solutions are as good as the prompts we give it,” he said. “If our prompts are not well designed, if our prompts are biased, then the outcomes of generative AI solutions will be biased.”
The Human Side
Using AI requires skill, Misra said, and companies should be upskilling and reskilling employees so that they attain AI fluency.
“What I mean by ‘AI fluency’ is, given a particular problem, how can a person, a new engineer or geoscientist, how can they decide what task can be delegated to AI? How will it be delegated to AI?” he said.
And, once the task has been delegated, he said, that employee must be able to identify appropriate and sensible outcomes.
Sander-Suicmez said companies should encourage and incentivize employees to learn how to use AI tools through taking courses and self-training.
And when a human is using AI well, El Bakry said, AI should expand the cognitive capability of the human.
“AI is not just about automation. It’s about amplifying the human’s capacity and capability and reimagining the future of energy,” he said.
