How Ready Are Universal Machine-learned Interatomic Potentials for Carbon Capture Simulations?
カーボンキャプチャーシミュレーションのためのユニバーサル機械学習原子間ポテンシャルの準備状況 (AI 翻訳)
Thiago Reschützegger, Guillaume Maurin, Cíntia Soares, Natan Padoin, Felipe Lopes
🤖 gxceed AI 要約
日本語
25種類のユニバーサル機械学習原子間ポテンシャル(u-MLIP)をCO2吸着予測にベンチマーク。9種のMOF実験データとの比較から、定量的予測は未だ困難であり、誤差源をMOF-CO2とCO2-CO2相互作用に分解する評価手法を提案。多様な材料スクリーニングへの道筋を示す。
English
Benchmarks 25 universal machine-learned interatomic potentials for CO2 adsorption in MOFs against experimental data, showing quantitative prediction remains out of reach. Introduces a protocol decomposing errors into MOF-CO2 and CO2-CO2 interactions, providing paths for reliable computational screening in carbon capture.
Unofficial AI-generated summary based on the public title and abstract. Not an official translation.
📝 gxceed 編集解説 — Why this matters
日本のGX文脈において
日本ではCCUSがGXの重要技術に位置付けられ、材料開発の効率化が急務。本論文はMLポテンシャルの精度限界を明確にし、今後の改良に具体的指針を与える点で、国内の材料スクリーニング研究に示唆を与える。
In the global GX context
Globally, computational screening is critical for accelerating carbon capture materials discovery. This paper provides the first systematic benchmark of universal ML potentials for adsorption, establishing an evaluation framework that can guide future method development across the CCUS community.
👥 読者別の含意
🔬研究者:Materials scientists and computational chemists should adopt the error decomposition protocol to improve ML potentials for adsorption.
🏢実務担当者:Companies engaged in carbon capture material discovery can use this benchmark to select appropriate simulation methods.
📄 Abstract（原文）
Nanoporous materials offer a tunable platform for selective carbon capture through adsorption, and identifying the best candidates requires computational screening built on accurate and transferable models of host–guest interactions. Classical force fields (FFs) have enabled rapid large-scale screening but fail to accurately describe such interactions for challenging systems, and while universal Machine-Learned Interatomic Potentials (u-MLIPs) promise to bridge this gap, their ability to predict macroscopic adsorption observables across chemically diverse frameworks remains largely underexplored. Here we benchmark 25 u-MLIPs against experimental CO2 adsorption data for nine metal–organic frameworks (MOFs) spanning rigid open-pore, open-metal-site, and confined small-pore classes, and find that quantitative isotherm prediction remains out of reach for the evaluated general-purpose models. We show that failures trace to two distinct error sources, the MOF–CO2 and the CO2–CO2 interactions, whose relative contribution shifts with framework chemistry, so that isotherm accuracy cannot be inferred from any single energy metric. In rigid open-pore and confined small-pore MOFs, the MOF–CO2 error dominates, while in strongly interacting frameworks with open metal site (OMS) both MOF–CO2 and CO2–CO2 descriptions become controlling factors, and error cancellation between the two sources can produce simulated accurate isotherms for physically incorrect reasons. Dataset composition, rather than architecture alone, determines whether a model reaches quantitative accuracy, and models trained on explicit MOF–CO2 and multi-molecule CO2 configurations perform most consistently. The evaluation protocol introduced here separates the two error sources and defines concrete paths for advancing u-MLIPs toward reliable computational screening for carbon capture.
🔗 Provenance — このレコードを発見したソース
- openalex https://doi.org/10.26434/chemrxiv.15006048/v1first seen 2026-07-16 05:22:01
🔔 こうした論文の新着を逃したくない方は キーワードアラート に登録（無料・3キーワードまで）。
gxceed は公開メタデータに基づく研究支援データセットです。要約・翻訳・解説は AI 支援で生成されています。 最終的な解釈・検証は利用者が原典資料に基づいて行うことを前提とします。