Özdemir and Yildirim (2025), Past, present and future of machine learning in chemical engineering

Seventeen pages and 117 references from Boğaziçi University, written as a history rather than a catalogue. The authors trace the field from the neural network work of the 1990s to large language models along three threads: algorithms giving way to an integrated workflow, small in-house datasets giving way to external databases, and applications widening from fault detection and process modeling to materials and property estimation.

The historical claim I found most clarifying is about naming. Neural networks, support vector machines and decision trees were in use in the field for decades before “machine learning” became the label; “data mining” was the common term in the early 2000s. The relabeling tracks a change in what a paper describes, from one algorithm to the whole chain of collection, storage and processing. A Web of Science keyword count, with data to July 2023, supports the claim that process and reactor design has been the largest application area in every decade, because control, identification, optimization and fault detection all fall into it.

The workflow section adds a step zero: decide the objective before choosing a method, because an unclear aim leads to a popular but unsuitable tool and results with no practical use. The rest is standard, with concrete details I appreciated: a missing pressure in a dataset very likely means atmospheric, a missing pretreatment probably means none, and the authors’ own group imputes missing values with a model trained on complete records. Smaller models relative to the number of points are described as more reliable.

Three application classes get one worked example each. Perovskite screening by gradient boosting over 230,808 candidates narrowed to 204 by a chain of filters and a final quantum check. The authors’ own CO2 methanation model, 4,049 points from 526 experiments in 100 papers, with random forest predicting conversion and the honest note that a bounded quantity is overpredicted near zero and underpredicted near one hundred percent. And monitoring and control, which they say may have been the first applications and remain a large share.

The outlook is the part that matters for process development. General materials databases built on molecular or quantum properties cannot carry chemical engineering problems, because surface area, porosity, dispersion and the operating conditions are as influential as the material. Field-specific databases are called for and immediately conceded to be hard: industry protects its data, standardization across laboratories is nearly impossible (the same nominal 300 W xenon lamp differs by brand), and industrial records are non-standard and incomplete by nature. On hybrid models the argument is stated and not demonstrated: a purely data-driven model has no knowledge of equilibrium and nothing stops it from predicting a conversion above it.

What to be careful with. It is a narrative review with no search protocol beyond the bibliometric figures; the 2020s bar in its keyword chart is four years scaled to ten; the examples lean toward catalysis and materials, which is the authors’ field; and the k-fold description on page 7 says the test set is divided for cross-validation where the training set is meant.

Read for the history, step zero and the database argument. The hybrid modeling section is a promise, not a result.