Al Sharah et al. (2024), Machine learning in chemical engineering, outlook and perspectives

Twelve pages, about seven of them text, 123 references, four schematic figures and no data. That is the shape of this review, and it tells most of the story. The authors, from three universities in Jordan, set out to judge whether machine learning in chemical engineering will stay a hot topic or fade, and they answer by listing what it offers and what it costs.

The frame is three linked components: data, representations and models. Data come from experiments, theory or simulation, and because experimental datasets are expensive the common route is an assembled database or text extraction from the literature. Representations are treated as a molecular question: SMILES and InChI strings, descriptors, Coulomb matrices, Morgan fingerprints, message-passing networks. Models are grouped by learning type, with the familiar list from principal components through random forests to recurrent networks. Process variables as inputs get one clause, and time series, historians and pilot units are absent.

The opportunities section is the more useful half. Trained models run fast enough for real-time optimization where first-principles models are too slow; a model can learn the discrepancy between a cheap estimate and the right answer; unsupervised methods find anomalies in streaming data; a learned surrogate can replace the slow inner model of a multiscale simulation. The authors propose an “as simple as reasonably possible” principle for model complexity, by analogy with risk management, and I have found myself quoting it.

The drawbacks section is where the paper earns a place on this shelf. Easy access lets people apply methods they do not understand. Hyperparameter tuning is trial and error at scale. A model can be right for the wrong reason. Performance collapses outside the training range, so the applicability domain must be stated. Positive results are published and negative ones are not, and engineers do not run processes at conditions they know to be bad, so the record lacks the failures a model would need in order to learn the limits. The last point is the one I think about most, because pilot campaigns are the purest case of it.

What to be careful with. There is no search method, no inclusion rule and no quantitative synthesis; coverage is the authors’ judgment. The argument follows Dobbelaere and colleagues (2021) closely, and several technical statements are imprecise. I use the paper to find its sources and to borrow its list of drawbacks, not to cite a number.

Read for the drawbacks list and the ALARP analogy. Then read the 2021 paper it leans on.