Investment bankers are speculating whether legacy businesses using artificial intelligence (AI) will yield productivity increases. Their thoughts are part of studies released this year by Stanford and MIT and first reported by Bloomberg. According to the studies, productivity has been flat for decades, but there’s hope that AI will catalyze an uptick.
The studies report AI can be a great benefit to novice, call center-support providers—it excels at learning from their scripts and fault trees. However, will we be as impressed when ChatGPT or its ilk are asked to assist with decisions in refineries or process plants?
Process plants running continuous or batch operations often lend themselves to mathematical modeling, and we’ve heard for years how simulations (digital twins) will help us increase yields and avoid unplanned outages. However, useful AI for a process plant probably needs more than procedures and rote knowledge to derive advice for a process plant operator. It must foresee the consequences of its advice.
Cause-and-effect inferred from process history is prone to link correlated measurements, rather than identifying root causes. Historical data can only show “where we’ve been.” Still, if AI increases on-stream time thanks to keeping the novice “between the lines” of established operational boundaries, it can be as valuable as throughput achieved via optimization. Nonetheless, a high-fidelity simulation running in parallel would be indispensable for vetting (and improving) AI advice.
Creating a high-fidelity simulation for even one operational unit can cost millions, not to mention its ongoing maintenance. The prospects of charming capital to fund controls projects—beyond rudimentary stay-in-business investments—are on the wane. For many of us, AI investments might have to be postponed in favor of “replacing the boiler from the 1960s before it breaks and shuts us down for a month.”
Big oil used to have the funds for such projects, but many of these downstream assets have been divested. When big oil retreats, capital goes with it. Salaries to attract talent likely to promote, justify and support complex investments in contemporary simulators, as well as instrumentation and controls, are uncommon. Meanwhile, we’re still “staffed to run,” so how will we find bold and determined (and capable) individuals to join our depleted ranks and lead us out of the doldrums as investors hope?
ISA 106 standard for Procedure Automation for Continuous Process Operations was approved this year, and perhaps its guidance can be enhanced by AI, provided an algorithm is created for it to capture, record, and digest good or bad outcomes. Perhaps AI can detect what delayed a startup and make suggestions to the next crew as they make their way through the procedure. Or, a crew member can leave a note saying, “The east reflux pump on the depropanizer cavitated due to condenser fouling.”
Will AI take a year to figure out that cooling water isn’t as cool on a humid summer day as it is in November? One can imagine that sensors for condition monitoring and seasonal changes will be necessary to better tune AI advice and learning. Procedure automation can also spur us to monitor the position of manual valves, which can allow AI-driven teams in remote control centers (possibly in another state) to coordinate more efficiently with operators in the field.
I find it hard to imagine effective AI that doesn’t have integrated sensors to “see” all the conditions that become part of an experienced operator’s decision-making process. How will AI know that water has condensed in the long-idle Vac Bottoms tank, and it will blow its lid and rain black oil on the community if it sends its 700 °F tower residual there? Will it know to check measurements that are flat-lining, or check whether the heat tracing has failed on a transmitter’s enclosure or impulse line (tubing)? All these added sensors will require a digital infrastructure, not only to bring in new sensors, but also to bridge data from rotating machinery diagnostics or electronic logbooks.
Using AI might add some sparkle to the next analysts call, and if the stock goes up, so might your promotion prospects. But all forays into emerging tech need a champion, and the champion may need to stay with his or her creation for a while. It must deliver value as promised. Operator advisory systems have been deployed for decades, and it only takes a few pieces of bad advice—or painfully obvious advice—for it to be ignored.
