Review in Chemical Society Reviews Maps Where AI Adds Value to Chemical Research, and Where It Falls Short

A review published in Chemical Society Reviews, one of the leading review journals in chemistry and a title of the UK-based Royal Society of Chemistry, sets out to answer a plain question: where does artificial intelligence actually help a working chemist, and how can anyone tell? The paper comes from researchers at the Skoltech AI Center and the Zelinsky Institute of Organic Chemistry of the Russian Academy of Sciences, with colleagues from ITMO University and other institutions.

The authors organize the review around the decisions a chemist makes rather than around the algorithms. Chemical research generates scientific publications, images, spectra, measurements and reaction records, and understanding how molecular structure relates to properties remains central. Along the way, researchers choose which evidence to examine, which experiment to run next, and how to change the plan once a new result arrives. The review treats that sequence as its framework.

Within it, the authors divide research into stages that generate, transform and validate information, then map AI methods onto each. They cover classical machine learning, models that build in chemical knowledge, and emerging foundation models for chemistry. The territory spans extracting knowledge from the literature and accumulated datasets; interpreting data from spectroscopy, mass spectrometry and microscopy; predicting properties and reactivity; planning syntheses; and systems in which one experiment's outcome guides the next. Representative applications show what transfers to a new problem and where a method's limits matter. The emphasis falls on measurable outcomes and recurring failure modes, not on benchmark scores alone.

In their conclusion, the authors name three ways AI can change a chemist's work. It can draw mechanistic insight from complex spectra, images and scientific texts. It can pick out promising compounds, synthesis routes and experimental conditions from among many options. And it can link predictions to automated experiments whose results steer the next step. One figure uses property prediction to show which steps between a molecular description and an experiment different models can handle. The practical value of AI, they argue, depends on whether it improves decisions under real research constraints.

The cautionary half of the paper concerns how success can be overstated. The authors examine information leaking between training and evaluation, overfitting to a limited set of examples, and errors when a model extrapolates beyond familiar data. These problems weigh heavily in chemistry, where substances, reaction conditions and instruments vary, and where experimental validation takes time and resources. Before adopting a model, they advise checking whether its training data match a particular laboratory's compounds and instruments, how reliably it estimates its own prediction uncertainty, and whether it has been tested on new experiments whose results were unknown during training.

Alexander Ryabov, Senior Research Scientist at the Skoltech AI Center, described the aim as viewing AI from the standpoint of a chemist who must make decisions. High accuracy on test data, he said, is only the tip of the iceberg; what matters more is whether a model helps interpret measurements, plan informative experiments, and keep predictions reliable on new compounds and reactions.

Michael G. Medvedev, head of the Theoretical Chemistry Group at the Zelinsky Institute and first author, looked ahead to a combination of advanced large language models and physics-informed models that respect chemical constraints. An AI agent, he said, must judge when a physics-informed prediction is well founded, when it goes beyond its training data, and which experiment would best reduce uncertainty. The researcher remains responsible for defining the scientific question and verifying what the agent produces.

The Skoltech co-authors include Ryabov; Vladimir Vanovskiy, Head of Hybrid Modeling; and RAS Professor Evgeny Burnaev, director of the AI Center. Zelinsky Institute co-authors include Medvedev and RAS Academician Valentine Ananikov. Burnaev, Medvedev and Ananikov are Sber Science Prize laureates. Looking forward, the authors argue that chemistry-grounded foundation models and intelligent experimental platforms should be evaluated across full research cycles, from the initial question to a reproducible result.