Yokogawa and JSR say they have completed a field test in which software ran a chemical plant autonomously for 35 consecutive days, a result the two companies describe as a world first. The test was carried out at a JSR plant and used reinforcement-learning software to control operations that ordinarily need an operator to adjust valves by hand.

The problem it addresses is specific. Established control methods, such as PID control and advanced process control, handle steady conditions well but struggle when something changes suddenly. A shift in air temperature caused by rain or snow is enough to push a distillation column out of balance, and the usual answer is for an experienced operator to take over and reset the control values.

What 35 days of autonomy proves

During the trial, the software kept product quality within specification and held the liquid level in the distillation column, while making use of waste heat as a heat source. Weather did disturb conditions, but the products made during the run met the standards and were shipped, which is the part that gives the result weight: the output was sold, not scrapped.

The algorithm, Factorial Kernel Dynamic Policy Programming, was developed by Yokogawa and the Nara Institute of Science and Technology in 2018. It had been exercised in simulations and in a training system before this plant trial. Masataka Masutani, JSR's general manager of production technology, and Takamitsu Matsubara of the institute both speak of the technology as answering a shortage of experienced operators as much as an efficiency problem.

One plant, for 35 days, is a demonstration. Whether the approach holds across other processes, and across the unplanned disturbances that a real year brings, is the question the two companies say they will now investigate.