Cloud synthesis: a global close-loop feedback powered by autonomous AI-driven catalyst design agent

MAINTEXT

Artificial intelligence (AI) is driving revolutionary advancements in the field of catalysis science, particularly excelling in catalyst property prediction, composition optimization, and the discovery of novel catalysts. Data-driven approaches, through the analysis of extensive datasets, have accelerated the automation of catalyst design and experimentation. However, current methods face several limitations. First, predictive models in the early stages often rely excessively on a single category of algorithms, such as machine learning regression models or large language models. This reliance compromises both interpretability and predictive accuracy, rendering the current automated exploration of new materials fundamentally a trial-and-error process. Second, effective catalyst design at the front end depends on diverse catalyst material databases. However, individual research groups are typically limited in the range of catalysts they can synthesize and provide feedback.

To address these challenges, we propose an AI-driven, large language model (LLM)-integrated automated design workflow based on the Digital Catalysis Platform (DigCat: https://www.digcat.org)^[1]. This workflow combines over 400,000 experimental performance data, more than 400,000 catalyst structures, state-of-the-art machine learning models, and catalytic reaction microkinetic models. Furthermore, by deploying this platform as a cloud-based predictive tool accessible via the Internet, it provides researchers worldwide with a cloud-based prediction and data feedback interface, fostering collaboration and advancing catalysis research globally.

As illustrated in Figure 1, catalyst researchers around the world can access the cloud-based design system and initiate a query as simple as: “Please design a new catalyst.” The DigCat online platform then leverages its repository of knowledge and models to execute an autonomous workflow for catalyst design.

Figure 1. The Digital Catalysis Platform (DigCat) serves as a cloud-based system for synthesis and global closed-loop feedback, powered by an autonomous AI-driven Catalyst Design Agent. AI: Artificial intelligence.

This design process involves five key steps. First, the system utilizes LLMs combined with experimental performance and structural databases of catalytic materials [Figure 2A and B]. By analyzing existing material compositions and structures, it generates potential material structure and composition schemes. In addition to analyzing catalyst-specific databases, the platform also integrates databases from other material domains, such as solid electrolytes, thermoelectric materials, and lithium battery electrode materials. By leveraging stability and fundamental performance insights from these broader material databases, the platform significantly expands the scope of potential candidates for catalyst material development. This comprehensive approach ensures that the catalyst design process benefits from cross-disciplinary data and insights.

Figure 2. Workflow of the cloud synthesis and global feedback system: (A) The interface of the front-end DigCat website; (B) Key tools within DigCat, including the experimental and theoretical structure databases, RGA-enhanced LLM, surface Pourbaix diagram and aqueous stability, machine learning regression model, catalytic microkinetic model, and the active machine learning training framework; (C) CPU and GPU infrastructure supporting DigCat; (D) Automated synthesis platforms worldwide, with images from Tohoku University in Japan and Beijing University of Chemical Technology in China. LLM: Large language model; DigCat: digital catalysis platform; CPU: central processing unit; GPU: graphics processing unit.

The second step involves evaluating the fundamental stability and cost of catalyst materials to enhance the likelihood of their practical application in the future. Stability analysis includes surface Pourbaix diagram analysis (CatMath^[2,3]) and Aqueous Stability Assessment^[4], both of which have been integrated into the DigCat online platform and can be readily executed by the system [Figure 2B]. The third step employs machine learning regression models to predict the adsorption energy and adsorption activity of catalysts. In this aspect, we have pre-developed code available to support the automated workflow, which can be accessed via the Machine Learning Modeling module in DigCat. Based on the adsorption energies predicted by the machine learning regression models, we use traditional thermodynamic volcano plot models to screen potential catalysts. These candidates are further evaluated using adsorption energy predictions based on machine learning force fields. To support this, we have developed a methodology integrating molecular dynamics and time-stamped force-biased Monte Carlo methods, enhanced by the Gaussian Approximation Potential^[5]. This active machine learning model effectively uncovers the microscopic processes involved in substrate-catalyzed growth, providing deeper insights into catalytic performance and behavior [Figure 2B].

Finally, the automated front-end DigCat platform incorporates the selected candidate catalysts into our state-of-the-art pH-dependent microkinetic models for reactions such as the oxygen reduction reaction (ORR)^[6], oxygen evolution reaction (OER)^[7], hydrogen evolution reaction, carbon dioxide reduction reaction (CO₂RR)^[8], and electrocatalytic ammonia synthesis. These models comprehensively account for the effects of electric field-pH coupling, thermodynamic energies, kinetic barriers, zero-point energy and entropy, solvation effects, and constant potential corrections. Furthermore, these models have been experimentally validated, demonstrating excellent agreement with experimental data.

How the closed loop is established after a candidate is discovered, synthesized, and its data reintegrated for improved predictions is also an important point for clarifying the workflow. The closed loop in our platform is established through iterative integration of AI-driven design [Figure 2C], automated synthesis [Figure 2D], experimental validation, and continuous feedback. As shown in Figure 3, the process begins with the DigCat, where the Catalysis AI agent proposes new catalyst candidates (defined by elements, structure, active sites, and synthesis routes) using a combination of machine learning and physical models. These candidates are then automatically synthesized and tested with high-throughput experimental devices, where both performance and characterization data are collected.

Figure 3. Schematic of the closed-loop DigCat, where AI-driven catalyst design is integrated with automatic synthesis, high-throughput experiments, and feedback to continuously improve catalyst discovery. DigCat: Digital catalysis platform; AI: artificial intelligence.

If the experimental outcomes confirm the AI-designed predictions, the candidate can be validated as a successful discovery. If not, the discrepancy between predicted and experimental results is explicitly analyzed. The experimental data are fed back into the platform, allowing the AI agent to update both machine learning models and underlying physical models with new physical insights. This adaptive learning step improves the accuracy of subsequent predictions.

The closed loop is not limited to one design-test attempt; it keeps learning from experimental feedback and improves with every cycle. Each round of experimental results enriches the database and refines the AI agent’s predictive capability, ensuring that future catalyst proposals become increasingly reliable and efficient. Ultimately, the closed-loop platform achieves the final objective of accelerating and improving catalyst discovery by coupling autonomous experimentation with self-improving AI models.

CONCLUSION AND OUTLOOK

At the forefront of innovation, the DigCat is built around the core principles of cloud-based synthesis and global closed-loop feedback, enabling a transformative approach to catalyst development. By connecting researchers through a unified digital framework, the platform seamlessly integrates computational predictions, automated workflows, and experimental feedback. This closed-loop system not only accelerates catalyst discovery but also ensures continuous optimization and refinement.

Looking to the future, DigCat aims to expand its capabilities through collaborations with automated synthesis platforms in multiple countries. These partnerships will establish a truly global network for closed-loop catalyst development, combining cloud-based tools with state-of-the-art synthesis and feedback mechanisms. This vision promises to unlock new possibilities in catalysis research, driving advancements in energy, sustainability, and environmental technologies.