CatDT
Self-evolving multi-agent system that constructs an autonomous, condition-aware digital twin of a working heterogeneous catalyst from only a bulk crystal and a natural-language reaction description.
| Affiliation | HKUST (IAS Center for AI for Scientific Discoveries; Departments of Chemistry, CSE, and Chemical & Biological Engineering) |
| First introduced | 2026-06 (arXiv preprint) |
| Lifecycle stages | Multi-stage (facet weighting, surface reconstruction, pathway enumeration, transition-state location, microkinetics) |
| Autonomy level | Semi-autonomous (5–30 min single-GPU runs from bulk crystal + reaction description; no manual intervention at intermediate links) |
| Domain focus | Chemistry / heterogeneous catalysis (gas–solid and liquid–solid) |
| Availability | Code on request — repository announced for release at github.com/AI4QC/catdt upon publication |
Approach
Eight specialized agents coordinate 27 scientific tools to predict stable facets, reconstruct surfaces under operating conditions, enumerate and rank reaction pathways, locate transition states, and compute kinetics. Two innovations address the system’s hardest steps. UniMech identifies dominant pathways for novel materials and under-explored reactions by fusing agent-guided proposals with energy-cached graph search over autonomously constructed reaction networks, operating at over 10³× lower cost than exhaustive enumeration. A memory-augmented reinforcement loop lifts barrier-calculation success from 41% to 84% across 600 diverse catalytic surfaces by learning to construct better initial and final states and path interpolations across runs.
The authors frame the decisive factor for a faithful catalyst digital twin not as raw LLM capability but as the engineered harness around it: deterministic tools, persistent memory, and verified self-improvement that compound across foundation-model upgrades, new tools, and runs.
Validation
Seven gas–solid benchmarks: stepped metals, single-atom catalysts, ordered intermetallics, vacancy-rich 2D sulfides, 2D carbides, and a strong-metal–support-interaction (SMSI) interface. Every CatDT prediction lies between 0.5× and 2× experimental values across measurements spanning four orders of magnitude.
Notable results
- Barrier-calculation success across 600 diverse catalytic surfaces improved from 41% to 84% via the memory-augmented reinforcement loop.
- UniMech finds dominant pathways at >10³× lower cost than exhaustive enumeration.
- For propane dehydrogenation, CatDT independently discovered non-precious catalyst candidates rivaling the Pt-based industrial benchmark; the proposed Ni@ZrO₂ SMSI overlayer reached a simulated TOF of 1.63 s⁻¹ at ~100% selectivity.
Primary paper
Other references
None.
Code
Announced release at github.com/AI4QC/catdt upon publication. RL-training trajectories, a 569-species thermodynamic-correction database, and CP-aware pretrained model parameters announced for release on Hugging Face upon publication.