ActMOF: Mapping Reticular Synthesis through Active Learning
Authors: Hongrui Sheng, Shutian Liang, Reid A. Coyle, Shyam Chand Pal, Peter Walther, Bin Feng, Xinyu Ding, Zhiling Zheng
Organizations: Department of Chemistry, Washington University, St. Louis, Missouri 63130, United States · Department of Chemistry, Fudan University, Shanghai 200438, China · Department of Energy, Environmental and Chemical Engineering, Washington University, St. Louis, Missouri 63130, United States · Institute of Materials Science and Engineering, Washington University, St. Louis, Missouri 63130, United States · Department of Chemistry, Imperial College London, London W12 0BZ, United Kingdom
Reticular chemistry allows metal-organic frameworks (MOFs) to be designed from their building units, yet their synthesis remains constrained by sparse experimental data, noisy crystallization outcomes, and the absence of benchmarks for comparing active-learning strategies. While data-driven approaches such as Bayesian optimization (BO) can improve synthesis success rates, selecting appropriate optimizers beforehand for mapping MOF crystallization landscapes remains largely empirical. Here we present ActMOF, an experiment-grounded benchmark for comparing search strategies in aluminum-carboxylate MOF synthesis. ActMOF is emulated from experimentally measured microwave syntheses with each encoded condition assigned a crystallinity index calculated based on diffraction peak intensity and full width at half-maximum. By benchmarking policies guided by Gaussian processes (GP), tree ensembles and large language models (LLM), ActMOF placed the GP family among the leading policies for active learning in sparse synthesis landscapes. It then demonstrated transferability by guiding the microwave synthesis of six reticularly related aluminum MOFs. All six campaigns gave crystalline products within a fixed budget of 12 experiments each, and established a reusable testbed for AI-driven MOF synthesis.
Figures & tables
Figure 1: Overview of the ActMOF active-learning benchmark. (a) Closed-loop active-learning workflow, in which proposed conditions are evaluated on the synthesis landscape (dry lab) or by microwave synthesis and PXRD (wet lab). (b) Structure of MOF-321, [Al(OH)(PZVDC)], with its linker H 2 PZVDC and rod-shaped Al-SBU. (c) The 95 measured MOF-321 syntheses across the five encoded reaction parameters, colored by experimental q . (d) Definition of q as peak height over full width at half-maximum (FWHM) of the diagnostic reflection, shown on a representative PXRD pattern. For illustrative clarity: Al is represented by deep blue octahedra; O, pink; C, light gray; N, green. H atoms are omitted.
Figure 2: Synthesis landscape and parameter distributions across q tiers. (a) Uniform Manifold Approximation and Projection (UMAP) of a 1% stratified random subsample of ∼ 6.1M enumerated synthesis conditions, colored by q percentile tier. Zero-crystallinity conditions ( q = 0) are shaded by temperature–metal quadrant, split at the dataset medians: T < 100 °C (Low T) or ≥ 100 °C (High T), and metal < 1.66 equiv (Low Metal) or ≥ 1.66 equiv (High Metal). Blue points mark the 95 experimental ground-truth conditions. (b) Ridgeline kernel density estimates (KDE) of each q tier along the five synthesis parameters, from a 2% random subsample within each tier; each curve is normalized to its maximum. The favorable window narrows with increasing tier stringency. Metal and modulator equivalents are mmol of AlCl 3 ·6H 2 O and NaOH, respectively, per 1 mmol of linker; added solvent is deionized water. UMAP and KDE settings are given in Supporting Information, Section S2.6 .
Figure 3: Benchmark comparison of active-learning policies on the ActMOF landscape. Cumulative best q is plotted against iteration (three random starting experiments, then 39 batches of three; 120 experiments in total) and averaged over five runs per policy, each run starting from its own random experiments. Shaded bands are 95% confidence intervals of the mean for random search and Matérn-5/2 PI. Shown are 16 of the 39 policies: Gaussian processes (Matérn-5/2 or Matérn-3/2 kernel; PI or EI), GP–LLM hybrids, tree ensembles (conformal random forest; Bayesian additive regression tree (BART)-like ExtraTrees), local approximate GPs (laGP), linear bandits (LinUCB), LLM-only proposers (GPT-5.5, GPT-4o, GPT-5-mini) and random search (Supporting Information, Table S13 ). The right axis normalizes each curve to the highest plotted q . Inset: iterations 0–5 for Matérn-5/2 PI and EI.
Figure 4: Selected experimental conditions and cumulative best q for a Matérn-5/2 Gaussian process with expected improvement (batch size = 3; mean of five runs). Heatmap rows correspond to the five decision variables (metal, modulator, additional solvent, reaction time, temperature); color encodes the normalized selected value at each iteration. The pink curve with right axis shows best q so far; the shaded band is the 95% confidence interval of the mean.
Figure 5: PXRD patterns for the six MOF systems used to validate transfer of the benchmark-guided active-learning policy (CAU-10, CAU-23, MIL-53-muc, MIL-160, MTV-MIL-160, and MTV-MOF-303). Crystal structures are shown above each column. For every system, representative experimental patterns from low-, medium-, and high- q outcomes encountered during the search are compared with the corresponding simulated reference pattern shown at the bottom of each column. Selected reflections in the high- q traces are indexed. For illustrative clarity: Al is represented by deep blue octahedra; O, pink; C, light gray; N, green; S, yellow. H atoms are omitted.
Chemical
M (g mol -1 )
Use
Supplier
Purity
Table S1 : Chemicals used for the validation syntheses, with molar masses M . Deionized (DI) water was used throughout.
Scalable synthesis remains the gate between MOF discovery and industrial deployment, as scale-up know-how is fragmented across disparate reports. We introduce ScaleMOF, a literature-mined dataset and a positive-unlabeled learning strategy that fine-tunes large language models. Achieving 93.5% accuracy, this proof-of-concept serves as a literature-grounded ranking tool prioritizing plausible scale-up candidates.
Peter Walther, Hongrui Sheng, Xinxin Liu +7
Department of Chemistry, Washington University, St. Louis, MO 63130, USA. · Department of Chemistry, Fudan University, Shanghai 200438, China. · Department of Computer and Information Science, University of Pennsylvania, PA 19104, USA. +2
Inverse design of metal-organic frameworks (MOFs) requires searching a combinatorially vast space where property labels are expensive and most machine-learning models reveal little about why a structure succeeds. We introduce LLM4MOF, a closed-loop framework in which language-model agents reason about chemistry, build candidate MOFs, and test them in simulation, refining hypotheses over ten autonomous iterations. One agent proposes interpretable design hypotheses over metal nodes, linkers, pore geometry, and functional chemistry, and a second translates them into constraints that select candidate MOFs, each made of a metal node, organic linker, and matching topology. Each hypothesis is tested through four diagnostic beams that apply different subsets of its constraints, so comparing them shows whether geometry, chemistry, or metal choice drives performance. Even when blind to the global property landscape of databases, LLM4MOF concentrates its search on top-performing structures across six adsorption, separation, and electronic-structure tasks within 400 property evaluations. The same loop also generates new MOFs de novo and validates them in live simulation, where it adapts the geometry to each requested condition, outperforming random search and a genetic algorithm at roughly $1 per campaign. LLM4MOF shows that language-model agents can run interpretable, simulation-grounded inverse design without training a model per objective.
Kyungmin Nam, Seunghee Han, Jihan Kim
Department of Chemical and Biomolecular Engineering, Korea Advanced Institute of Science and Technology, Daejeon 34141, Republic of Korea.
Metal-organic frameworks (MOFs) offer a vast design space, and as such, computational simulations play a critical role in predicting their structural and physicochemical properties. However, MOF simulations remain difficult to access because reliable analysis require expert decisions for workflow construction, parameter selection, tool interoperability, and the preparation of computational ready structures. Here, we introduce SimMOF, a large language model based multi agent framework that automates end-to-end MOF simulation workflows from natural language queries. SimMOF translates user requests into dependency aware plans, generates runnable inputs, orchestrates multiple agents to execute simulations, and summarizes results with analysis aligned to the user query. Through representative case studies, we show that SimMOF enables adaptive and cognitively autonomous workflows that reflect the iterative and decision driven behavior of human researchers and as such provides a scalable foundation for data driven MOF research.
Jaewoong Lee, Taeun Bae, Jihan Kim
Department of chemical and Biomolecular Engineering, Korea Advanced Institute of Science and Technology (KAIST), 291 Daehak-ro, Yuseong-gu, Daejeon 34141, Republic of Korea