Chowdhury et al. (2026), AI and machine learning in process systems engineering

The authors describe a five-step selection: 1,250 records from publisher databases, patents, white papers and websites, cut to 350 on abstract screening and 175 on full reading. That is the most method any of the reviews on this shelf reports, and it still gives no inclusion rule, no screening log and no list of what survived. The search drew on the open web as well as the journals, and the industrial sections show it.

The frame is that first-principles process systems engineering and data-driven methods are complementary. A Likert heat map scores the two approaches on scalability, adaptability, computational cost, explainability and implementation cost; the authors say the scores are their own reading of the literature and indicative only, which is honest and also means the figure carries no information beyond the text. The pairing of strengths is the useful version of the same idea: pattern recognition from noisy data beside rigorous modeling, real-time adaptation beside accuracy under known conditions, early fault detection beside mechanistic root cause analysis.

The catalogue is the reason to keep the paper. Tables 2 to 4 list reported outcomes by function: a hybrid ethylene oxide reactor model with 0.71 percent mean squared error against 5.99 percent for the mechanistic model; a crude unit energy efficiency raised from 51.9 to 70.6 percent by a neural network; a long short-term memory model inside predictive control with 46.3 percent better closed-loop stability; a reinforcement learning controller said to have run a distillation column for 35 consecutive days. I checked the tables against the page images and every value is the review’s report of someone else’s result. Each one has to be confirmed in the primary source before it is used, and in two cases the citation given does not support the statement it sits beside.

The barriers section holds the one sentence I cite most from this paper. Datasets dominated by steady-state operation give models that do not generalize to transients, and the decisions of past operators are embedded in the historical record. Pilot campaigns are steady state by design, so the sentence describes my data exactly. The rest of the list is familiar: formats that differ between the control system, the laboratory and the operator log; rare faults and therefore imbalanced fault data; models operators cannot trace and so will not act on; training cost; legacy control systems with no way in.

What to be careful with. About thirty of the 174 references are company or vendor web material, and the roll call of deployments is built from press releases. There is no pilot-scale or scale-up content at all. The paper is a prior-review entry and a map of what has been claimed, not a source of facts.

Read for the steady-state sentence and the barrier list. Use the tables as a reading list, never as data.