Autonomous heterogeneous catalyst discovery with a self-evolving multi-agent digital twin

Researchers introduce CatDT, a self-evolving multi-agent AI system that autonomously discovers heterogeneous catalysts by building digital twins of working catalytic systems. The system achieves predictions within 0.5-2x of experimental results across diverse catalyst types and independently identifies non-precious catalyst candidates for propane dehydrogenation that rival industrial platinum-based benchmarks.

CatDT represents a significant advancement in computational catalysis by addressing a persistent gap between theoretical predictions and experimental reality. The system combines eight specialized AI agents with 27 scientific tools to model complex catalytic processes—from surface reconstruction to kinetic calculations—in minutes rather than hours or days. This efficiency gain stems from two key innovations: UniMech, which reduces computational cost by over 1000x through intelligent pathway enumeration, and a memory-augmented reinforcement loop that improves transition-state calculation success rates from 41% to 84%.

The breakthrough demonstrates that faithful scientific simulation depends less on raw language model capability and more on engineering discipline—deterministic tools, persistent memory systems, and verified self-improvement mechanisms that compound across iterations. This architectural insight has broader implications for multi-stage scientific simulators beyond catalysis. By validating predictions across seven different catalyst benchmarks spanning metals, single-atom systems, intermetallics, and complex interfaces, CatDT establishes credibility for autonomous materials discovery.

For the catalysis and materials science communities, this work opens pathways to identify economically viable alternatives to precious metals. The discovery of Ni@ZrO₂ candidates achieving near-theoretical selectivity at competitive turnover frequencies suggests significant commercial potential. The practical impact extends to industries reliant on catalytic processes—petrochemicals, pharmaceuticals, and environmental remediation. Looking ahead, the validation of agent-driven discovery could accelerate adoption of autonomous systems in materials science, potentially reducing development timelines from years to months while lowering research costs substantially.

• →CatDT achieves experimental-level accuracy (0.5-2x) across diverse catalyst types using multi-agent AI orchestration
• →UniMech pathway discovery reduces computational cost by over 1000x compared to exhaustive enumeration methods
• →Memory-augmented reinforcement learning increased transition-state calculation success from 41% to 84% across 600 surfaces
• →System independently discovered Ni@ZrO₂ catalysts rivaling industrial platinum benchmarks for propane dehydrogenation
• →Engineered tooling and deterministic systems prove more critical than raw LLM capability for scientific simulation accuracy