INFORM April 2026

20 INFORM APRIL 2026 , VOL. 37, NO. 4

Song wanted to develop catalysts based on zeolites modified with common transition metals, which are less expensive and can operate under milder conditions than noble metal catalysts. So, he used an ML model to screen catalysts for the deoxygenation of guaiacol, a lignin‑derived compound abundant in bio‑oil. The reaction converts guaiacol into the high-value products benzene, toluene, and xylene (BTX). Song chose guaiacol as a model compound because its hydroxyl and methoxy groups make it particularly difficult to deoxygenate. The team selected ZSM-5 as the catalyst base because its pore structure and strong acidity enhance selectivity and activity. Song’s earlier work showed that doping ZSM-5 with metals can alter its catalytic performance. Using in-house and external datasets, the researchers identified 21 metals commonly used in deoxygenation, yielding over 1,330 possible three‑metal combinations. A Random Forest ML model screened and ranked these options. Although a ZSM-5 catalyst doped with zinc (Zn), gallium (Ga), and cerium (Ce) ranked fourth overall, it was chosen for validation because it contained no precious metals. The researchers synthesized the Zn‑Ga‑Ce/ZSM‑5 catalyst predicted by the model. Experiments showed that Zn and Ga improved methane activation at low temperatures, while Ce

reduced over‑coking, or the buildup of carbonaceous deposits. The optimized catalyst achieved guaiacol conversion with a liquid yield of 58 percent and BTX selectively as high as 90 percent— significant improvements over the unmodified ZSM-5 catalyst. In another study, Song’s team combined ML with process simulations to conduct techno-economic and life cycle assessments of biomass-to-biofuel conversion. An ANN model predicted how feedstock type, reaction temperature, and catalyst choice affected bio‑oil yield. These predictions were then fed into simulations to establish material and energy balances and to guide equipment design. The analysis showed that replacing hydrogen with methane in the deoxygenation step reduced the minimum selling price of renewable diesel by 22 percent and cut greenhouse gas emissions by 66 percent. The team notes that this AI-enabled workflow avoids the need for costly pilot plant optimization. Song is now collaborating with fuel companies, including Shell International, to test the methane-based process at a one‑ton‑per‑day demonstration scale. His team is also developing an AI chip that integrates optimization algorithms directly into facility hardware. “Companies like Shell deal with a wide variety of organic solid wastes,” Song says. “Our chip scans the feedstock and automatically adjusts pressure, temperature,

and flow rate to maximize production.” He emphasizes that the technology is still emerging and will benefit from further industry partnerships. Across the fats and oils sector, AI is moving from the lab bench to the factory floor. AI tools support yield prediction, catalyst and feedstock selection, equipment design, and the replacement of pilot‑scale testing. Digital twins and real‑time control systems are closing the loop between prediction and operation, enabling dynamic adjustments for efficiency, sustainability, and profitability. As sensor networks expand and models mature, the field is moving toward fully autonomous process control—systems that learn continuously, anticipate disruptions, and optimize entire value chains. Does this mean the process engineer’s days are numbered? Not necessarily. “AI can process millions of data points, but expert judgment is irreplaceable for interpreting results, adjusting models, and validating predictions,” says Mejuto. “In the olive oil industry, the goal is not to replace the master miller or tasting panel, but to give them real‑time, scalable decision‑making tools.” For hyperlinks to references visit inform.aocs.org Laura Cassiday is a freelance science writer and editor based in Hudson, Colorado. She can be reached at laura.cassiday.phd@gmail.com.