AI in Chemistry: Researchers Map Where Artificial Intelligence Delivers Value and Where It Falls Short

AI's Role in Chemical Research Workflows Examined

Researchers from the Skoltech AI Center, Zelinsky Institute of Organic Chemistry of the Russian Academy of Sciences, ITMO University, and other institutions have published a comprehensive review in Chemical Society Reviews examining where Artificial Intelligence produces measurable value in chemical research and where its reliability falls short1

1

. The review addresses a critical gap: why strong benchmark scores don't necessarily translate into laboratory performance, a concern that affects how AI changes chemical research in practice.

The authors organize their analysis around the decision sequence chemists follow during research, from examining evidence to running experiments and adjusting plans based on results. Chemical research generates diverse information including scientific publications, images, spectra, measurements, and reaction records, all of which AI methods in chemistry must process effectively2

2

.

Mapping AI Methods Across Research Stages

The review divides AI in chemical research workflows into stages that generate, transform, and validate information. The authors examine classical ML, models incorporating chemical knowledge, and emerging foundation models for chemistry, emphasizing measurable outcomes over benchmark scores alone1

1

. Coverage spans knowledge extraction from scientific literature, interpretation of experimental data from spectroscopy and mass spectrometry, predicting properties and reactivity, planning chemical syntheses, and systems where one experiment guides the next.

Alexander Ryabov, Senior Research Scientist at the Skoltech AI Center, explains the practical focus: "We wanted to look at AI from the perspective of a chemist who needs to make research decisions. A model's high accuracy on test data is only the tip of the iceberg. What matters more is whether it helps interpret measurements, plan informative experiments, and maintain the quality and reliability of its predictions when working with new compounds and reactions"2

2

.

Three Ways AI Transforms Chemistry Work

The authors identify three primary ways AI's impact on chemical research manifests. First, it extracts mechanistic insights from complex spectra, images, and scientific texts. Second, it identifies promising compounds, synthesis routes, and experimental conditions from numerous options. Third, it connects predictions to guiding automated experiments whose results inform the next step1

1

. AI's practical value depends on whether it improves decisions under real research constraints, not just theoretical performance.

Critical Risks and Reliability Concerns

The review examines risks that inflate apparent model success: information leakage between training and evaluation, overfitting to limited examples, and extrapolation errors beyond familiar data. These issues prove especially consequential in chemistry, where substances, reaction conditions, and instruments vary while experimental validation requires significant time and resources2

2

. Before deploying a model, researchers should verify whether training data match their laboratory's compounds and instruments, assess how reliably it estimates prediction uncertainty, and confirm testing on new experiments whose results weren't known during training.

Future Integration of LLMs and Physics-Informed Models

Michael G. Medvedev, Head of the Theoretical Chemistry Group at the Zelinsky Institute of Organic Chemistry and first author, outlines the next stage: "The next stage of AI in chemistry will bring together advanced large language models (LLMs) and physics-informed models that respect chemical constraints. An AI agent must assess when a physics-informed model's prediction is well founded, when it goes beyond the data used to train it, and which experiment would best reduce uncertainty. The researcher, in turn, remains responsible for defining the scientific question and verifying the results produced by the AI agent"1

1

. This integration represents where AI methods in chemistry must evolve to deliver consistent value across full research cycles, from initial questions to reproducible results.