LLaMalysis: Language-guided inverse design of catalyst surfaces with targeted hydrogen adsorption energies

Computational catalyst discovery is constrained by vast compositional and structural spaces, whereas atomistic generative models typically rely on engineered conditioning variables. Here we formulate catalyst inverse design as an instruction-following task in which compositional and catalytic requirements are expressed in natural language. We fine-tune an instruction-tuned language model on serialized catalyst slabs from the Open Catalyst 2020 dataset and obtain a generator with validity and uniqueness comparable to catalyst-specific baselines. Reward-weighted regression using machine-learning interatomic-potential feedback increases the number of candidates satisfying the validity criteria and lying within 0.10 eV of the target from 8 to a maximum of 65 per 500 generations. The refined model retains metal-specific controllability for ten metals, and density functional theory calculations of 101 prioritized candidates identify multiple alloy families with promising hydrogen adsorption energies. These results establish natural language as a practical conditioning interface for atomistic catalyst design.

Authors

Institutions

Publication Details

Journal
ChemRxiv
Published
2026-11-16
DOI
https://doi.org/10.26434/chemrxiv.15012306/v1
Primary Topic
Machine Learning in Materials Science
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

LLaMalysis: Language-guided inverse design of catalyst surfaces with targeted hydrogen adsorption energies

Tuan Anh Pham, Deahyun Kim, Haesun Park, Kwangnam Kim
ChemRxiv
Machine Learning in Materials Science
preprint

LLaMalysis: Language-guided inverse design of catalyst surfaces with targeted hydrogen adsorption energies

Tuan Anh Pham, Deahyun Kim, Haesun Park, Kwangnam Kim
preprint en

Abstract

Computational catalyst discovery is constrained by vast compositional and structural spaces, whereas atomistic generative models typically rely on engineered conditioning variables. Here we formulate catalyst inverse design as an instruction-following task in which compositional and catalytic requirements are expressed in natural language. We fine-tune an instruction-tuned language model on serialized catalyst slabs from the Open Catalyst 2020 dataset and obtain a generator with validity and uniqueness comparable to catalyst-specific baselines. Reward-weighted regression using machine-learning interatomic-potential feedback increases the number of candidates satisfying the validity criteria and lying within 0.10 eV of the target from 8 to a maximum of 65 per 500 generations. The refined model retains metal-specific controllability for ten metals, and density functional theory calculations of 101 prioritized candidates identify multiple alloy families with promising hydrogen adsorption energies. These results establish natural language as a practical conditioning interface for atomistic catalyst design.

ChemRxiv
Lawrence Livermore National Laboratory (US), Chung-Ang University (KR)
Machine Learning in Materials Science
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon