Artificial intelligence is transforming catalyst research by automating computational tasks that once consumed weeks of expert time, with new AI agents capable of completing molecular simulations and hypothesis generation in minutes rather than days. This acceleration could reshape how quickly new catalysts move from laboratory discovery to industrial deployment, though researchers warn that success depends on solving critical data standardization challenges.

What Are AI Agents Doing in Catalyst Research?

Catalyst research has historically been bottlenecked by the sheer volume of computational work required. Varinia Bernales, a computational chemist who previously worked at Dow Chemical, spent her days "creating files, submitting calculations, troubleshooting," she explained to Chemistry World. "There were more experimentalists than computational people. Those experimentalists have a lot of ideas to test, and unfortunately, I didn't have time to tackle them all." She would often validate hypotheses as weekend projects.

Now, AI agents like El Agente, which Bernales' team at the University of Toronto is developing in collaboration with Nvidia, are automating much of that work. These systems integrate different computational tools and can perform molecular simulations, review existing literature, and generate testable hypotheses without requiring researchers to write code. "With this tool, experimentalists will have another superpower," Bernales noted. "We need it because we need to make progress faster and make quality of life better".

The impact extends beyond individual labs. Ted Sargent from Northwestern University emphasized that the acceleration applies across the entire catalyst lifecycle. "It can take anywhere from a decade to 25 years to go from discovering a catalyst to using it in large scale applications," he said. "It's the whole life cycle that we want to accelerate".

How Are Machine Learning Models Speeding Up Simulations?

A key technological breakthrough enabling this acceleration is the development of machine learning interatomic potentials (MLIPs). These models, which have been refined since 2010, replace computationally expensive density functional theory (DFT) simulations with faster approximations trained on data. The speed improvement is dramatic: Fernanda Duarte of the University of Oxford noted that "we can do a calculation in 25 seconds that is perfect," compared to the hours or days required for traditional DFT methods.

MLIPs also enable simulations that were previously impossible, such as exploring how catalysts interact with complex solvent environments and molecular aggregates at higher concentrations. This matters because, as Duarte explained, "modifying one atom can eliminate [a catalyst's] effectiveness. It works perfectly, or nothing. In catalysts, aggregates form at higher concentration that we cannot predict when we compute single molecules. All these small details really matter".

However, the cost of running these simulations remains a concern. Graphics processing units (GPUs), which power the neural networks behind machine learning, are in short supply and expensive. Duarte noted that a GPU-powered computer she purchased three years ago costs more today. Researchers can access GPUs in remote data centers, but this adds expense. One cost-reduction strategy is training MLIPs on smaller, specialized datasets. Duarte's team has shown that models trained on just a few hundred DFT-simulated data points can achieve "highly accurate" results for specific chemical reactions.

Steps to Accelerate Catalyst Discovery with AI

• Standardize Data Collection: Establish consistent formats and metadata standards for recording experimental and computational catalyst data, enabling AI systems to learn from larger, higher-quality datasets across institutions and companies.
• Integrate Computational and Experimental Workflows: Connect AI simulation tools directly to laboratory automation systems, allowing AI agents to propose hypotheses and experimentalists to test them in rapid cycles without manual data transfer.
• Train Specialized Models: Develop machine learning models tailored to specific catalyst systems or reactions rather than relying solely on general chemistry models, improving accuracy and reducing computational costs.
• Expand Public Datasets: Contribute experimental and computational results to open-access initiatives like the Open Catalyst Project, Nomad, and the Online Materials Database (OQMD) to improve AI training and democratize access to computational tools.

What's Blocking Faster Progress?

Despite the promise, significant barriers remain. Graham Hutchings from the University of Cardiff, who heads the Max-Planck-Cardiff Centre on the Fundamentals of Heterogeneous Catalysis, emphasized that "it needs very clean data." His team is testing AI approaches on catalysts that convert carbon dioxide into useful substances, using models from the Nomad project that identify correlations between physicochemical properties. However, the lack of standardized data formats across institutions and industries limits what AI systems can learn.

The challenges include experimental validation of AI-predicted catalysts, scaling up promising laboratory results to industrial production, and filling gaps in available datasets. Many companies keep proprietary catalyst data private, limiting the information available for training public AI models. Additionally, the cost of both simulations and experiments constrains how many hypotheses researchers can test, even with AI acceleration.

How Does AI Fit Into Battery Materials Discovery?

The acceleration of materials discovery through AI extends beyond catalysts to battery chemistry, where the stakes are equally high. A comprehensive review published in June 2026 examined how AI and machine learning are transforming battery research across four critical areas: materials discovery, automated synthesis and testing, early lifetime prediction, and recycling.

Traditional research approaches struggle with the vast, high-dimensional search spaces involved in optimizing next-generation battery chemistries beyond conventional lithium-ion batteries. Data-driven methods using AI and machine learning offer promising routes to accelerate discovery and scale validation. The review identified several promising directions, including domain-specific foundation models, hybrid physics-aware approaches, and agentic orchestration of hardware and software to enable faster, safer, and more sustainable deployment of next-generation batteries.

Researchers are already applying these techniques. For example, studies have used machine learning to predict battery cycle life before capacity degradation occurs, enabling researchers to identify promising candidates early. Other work has employed AI to guide the design of electrolytes and solid-state conductors, areas where the chemical space is too large for traditional screening methods.

Why Does Data Quality Matter More Than Speed?

As AI tools become more powerful, the limiting factor shifts from computational capacity to data quality and availability. Researchers at major tech companies including Meta, Google, and Nvidia are confident that AI is well-suited to solving commercially and societally useful problems with catalysts and materials. However, they acknowledge that progress depends on access to larger, better-standardized datasets.

The best datasets for AI modeling in chemistry remain limited. While initiatives like the Protein Data Bank provide valuable resources, catalyst and materials research lacks equivalent centralized repositories with consistent data standards. This creates a paradox: AI can accelerate discovery faster than ever before, but only if researchers can train it on sufficiently large and clean datasets. Solving this data challenge may ultimately determine whether AI's promise to compress 10-25 year discovery timelines into months becomes reality.