AI Meets Catalysis: A New Scientific Paradigm Is Emerging

Artificial intelligence has swept through nearly every corner of modern science, from predicting protein structures to designing new materials, but one of chemistry’s oldest and most industrially vital disciplines has been surprisingly slow to join the revolution. A new open-access review published in Discover Chemistry by researchers at the Venezuelan Institute for Scientific Research and the Central University of Venezuela argues that catalysis stands at the threshold of a genuine paradigm shift, one the authors call the “Catalysis–AI” convergence. Yet their sweeping bibliometric analysis, spanning more than three decades of scientific literature, reveals a striking paradox: while both fields are booming individually, their intersection remains remarkably thin, accounting for a mere fraction of the combined output.

The numbers tell the story with unusual clarity. The Lens platform, which indexes scholarly publications across disciplines, lists more than 2.4 million papers related to catalysis and over 552,000 related to artificial intelligence. At the intersection of the two, the researchers found only 618 records, roughly 0.02 percent of the combined total. That figure is all the more remarkable given what happened next: since 2019, publications at the interface have grown at an annual rate of 18.34 percent, and output since 2020 has surged to 29 times its 2019 level. The earliest paper bridging the fields appeared in 1991, a fuzzy temporal model for fault diagnosis in a fluidized catalytic cracking unit, but for nearly three decades afterward the field remained essentially stagnant before exploding into exponential growth.

To understand why this convergence matters, it helps to recall what catalysis actually does. Catalysts accelerate chemical reactions without being consumed, underpinning everything from fertilizer production and petroleum refining to pharmaceuticals and emissions control. Traditionally, discovering and optimizing them has been an exercise in painstaking trial and error, guided by intuition and decades of accumulated expertise. The review traces the intellectual lineage of the computational alternative back to Alan Turing’s 1950 question of whether machines can think and John McCarthy’s 1956 definition of artificial intelligence at the Dartmouth Conference. It also revisits the expert systems of the 1980s, such as Carnegie Mellon’s XCON, which by 1986 had processed around 80,000 orders for Digital Equipment Corporation with nearly 98 percent accuracy and saved the company an estimated 25 million dollars annually, and Stanford’s MYCIN, which diagnosed infectious diseases with about 65 percent success compared with roughly 80 percent for human specialists.

Those early systems exposed the limitations that would define AI’s first decades: inflexibility, high maintenance costs, and thorny questions of responsibility when things went wrong. The turning point came with the backpropagation algorithm, advanced by Geoffrey Hinton and colleagues in the 1980s, which underpins virtually all contemporary neural networks. Only after 2006 did such networks begin to exert substantial influence, and the past five years have brought text mining of large-scale databases, natural language processing for structure–activity relationships, large language models, and the AlphaFold systems that predict protein three-dimensional structures and protein–ligand interactions from amino acid sequences. Multimodal AI platforms, autonomous agents, and self-driving laboratories have since begun to accelerate experimental workflows in ways that were computationally prohibitive only a few years ago.

One of the review’s most intriguing threads is the connection to mesoscience, a framework developed by Chinese researchers to describe the multiscale spatiotemporal structures of complex systems, bridging the gap between the macroscale behavior of a system as a whole and the microscale behavior of its constituent units. The Energy Minimization Multiscale Model, formulated as a multi-objective variational problem, has historically demanded massive processing resources. Physics-Informed Neural Networks offer a disruptive workaround: rather than acting as mere universal approximators, they embed fundamental physical laws directly into the network’s loss function, ensuring consistency with conservation principles and dramatically accelerating convergence. Applied to catalysis, this approach has already yielded concrete results. Peng and Wei showed how mesoscale analysis of electrode morphology, encompassing pore structure, reaction interface, and active site, can guide the design of more efficient electrocatalytic materials, while Wu and colleagues identified the heat of reaction and enthalpy of fusion as key factors governing stability in supported nanometallic catalysts for hydrogen peroxide synthesis.

Machine learning’s contribution to catalyst discovery is perhaps best illustrated by a landmark study from Ulissi and co-workers, who built an automated framework for exploring bimetallic catalysts. Their systematic search uncovered an active site that had previously been overlooked in nickel–gallium systems during the electrochemical reduction of carbon dioxide: individual nickel atoms encircled by gallium atoms at the surface. These sites displayed enhanced thermodynamic parameters and distinctive step-like kinetic behavior, substantially advancing the rational design of bimetallic catalysts. The lesson is a powerful one, the review suggests: algorithms trained on known catalytic properties can systematically explore chemical spaces too vast for any human team, surfacing candidate structures and mechanisms that intuition alone would never reach.

The bibliometric mapping also charts how the vocabulary of the field has evolved. Between 2001 and 2006, trending terms clustered around fluid catalytic cracking units, process variables, and petroleum engineering, reflecting the energy industry’s grip on early applications. From 2009 to 2017, computational tools such as genetic algorithms, fuzzy logic, and optimization algorithms consolidated. Then came the inflection: in 2021, protein engineering, catalyst discovery, and catalytic reaction became trending topics; by 2022, neural networks ranked as the third most frequent term in the corpus; and in 2023, artificial intelligence, machine learning, and deep learning consolidated as the dominant nodes. By 2024, the spotlight had shifted again, toward solar cells, energy materials, and reaction conditions, signaling a growing emphasis on renewable energy and efficient materials design.

Yet the density maps reveal an uncomfortable truth. The term clusters split cleanly into three subspaces, two dominated by artificial intelligence and one by catalysis, and the persistent physical distance between them in the co-occurrence maps confirms that Catalysis–AI is an emerging frontier rather than a consolidated discipline. The thematic evolution tells a subtler story: the AI cluster began the study period as a driving theme, transitioned through a basic, cross-cutting phase, and now sits in the quadrant of emerging or declining topics, though the authors interpret this not as decline but as internal reconfiguration, given that the final period alone produced 473 publications, 76.54 percent of the entire corpus, while the cluster structure consolidated from nine groups down to three.

The geographic distribution adds a sobering dimension. The United States and China each account for 17.80 percent of publications at the interface, followed by Germany, the United Kingdom, and India, with the top ten countries producing 64.40 percent of all documents. Within the Ibero-American community, despite decades of distinguished catalysis research, only seven of twenty-two countries have published in this field at all, contributing just 3.72 percent of global output, led by Spain at 1.62 percent, Brazil at 1.13 percent, and Mexico at 0.32 percent. The authors argue that bridging this digital divide is a strategic priority, and they point to low-computational-cost, open-source algorithms as a viable path for experimental groups that cannot afford supercomputing infrastructure. They call for regional data-science consortia, democratized access to open databases, and the early integration of chemometrics and machine learning into chemistry curricula to cultivate a generation of scientists fluent in both languages.

Perhaps the most provocative recommendation concerns data culture itself. The review identifies a lack of standardization in experimental reporting as a critical barrier to dataset interoperability, and it takes aim at publication bias: the tendency to report only successes and optimized yields. For AI models to achieve genuine predictive capability, the authors contend, the community must begin documenting negative results and failed reactions, because these provide the contrast needed to map reactivity boundaries and prevent overfitting. An AI that learns only from successes, they warn, is blind to the inherent complexity of the catalytic design space. The challenges that remain are formidable, from heterogeneous systems whose interconnected parameters evolve throughout a reaction to practical questions of metal cost, abundance, and scale-up. But the trajectory is unmistakable, and the authors suggest the coming decades may deliver a genuine human–machine synergy in the design of novel catalytic materials, with the 2025–2026 literature already being compiled for a follow-up study to track this rapidly evolving digital landscape.

Subject of Research: The integration of artificial intelligence and machine learning into catalysis research, assessed through a bibliometric analysis of publications from 1991 to 2024

Article Title: The emergence of the catalysis artificial intelligence paradigm transforms modern chemical and catalytic research

Article References: Villanueva, S., Mendoza, L., Betancourt, P., & Pinto-Castilla, S. (2026). The emergence of the catalysis artificial intelligence paradigm transforms modern chemical and catalytic research. Discover Chemistry, 3(1), Article 496. https://doi.org/10.1007/s44371-026-00926-9

Image Credits: AI Generated

DOI: 10.1007/s44371-026-00926-9

Keywords: artificial intelligence, catalysis, machine learning, bibliometrics, mesoscience, neural networks, catalyst design, electrocatalysis, protein engineering, Industry 4.0, Ibero-America, sustainable chemistry