PaperlayerView exact revision on arXiv

Source-linked AI summary

ActMOF: Mapping Reticular Synthesis through Active Learning

Hongrui Sheng, Shutian Liang, Reid A. Coyle, Shyam Chand Pal, Peter Walther, Bin Feng, Xinyu Ding, Zhiling Zheng

arXiv:2610.19404v1cond-mat.mtrl-scics.LGphysics.chem-ph

TL;DR

Finding crystalline MOF synthesis conditions remains challenging amid many reaction variables and narrow crystallization windows. ActMOF benchmarks search policies on an experiment-grounded landscape and applies benchmark-guided strategies to related frameworks; reported results include a 2.2-fold average increase ratio across six frameworks.

  • Problem

    Finding conditions for the intended crystalline MOF phase is difficult because crystallization depends on many reaction variables and small condition changes can produce noncrystalline outcomes.

    Source: Introduction · pp. 1–2

  • Method

    ActMOF uses a landscape emulated from measured syntheses and scores conditions with q, the diagnostic PXRD peak intensity divided by its full width at half-maximum.

    Source: Results and Discussion · pp. 3–4

  • Results

    2.2-fold average increase ratio across six frameworks; CAU-23 never exceeded its starting best q.

    Source: Results and Discussion · p. 8

  • Takeaways & Limitations

    Benchmark-selected active learning located linker-specific conditions for six targets when synthesis variables retained comparable chemical meaning.

    Source: Results and Discussion · p. 9

    ActMOF is a qualitative experiment-grounded landscape for comparing search strategies, not a predictor of every unmeasured condition’s experimental outcome.

    Source: Results and Discussion · p. 4

Abstract

from arXiv · show

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.

Introduction

MOF structures can be designed from building units, but finding conditions that yield the intended crystalline phase remains difficult because many variables shape crystallization. ActMOF benchmarks active-learning strategies on an experiment-grounded synthesis landscape and transfers the selected strategies to related aluminum MOFs.

Introduction

  • Crystallization depends on many interacting conditions, including concentration, modulator, solvent, time, temperature, and nucleation kinetics, making synthesis navigation increasingly pressing as reticular space expands.Source: Introduction · p. 1
  • Prior condition searches have relied on chemical intuition, literature analogy, and design-of-experiments campaigns.Source: Introduction · p. 2
  • A designed MOF may still crystallize only within a narrow window, where small changes in acidity, dilution, or thermal dose can yield amorphous solid, side phase, or unreacted precursor.Source: Introduction · p. 2
  • Active learning iteratively proposes, evaluates, and updates synthesis conditions using prior outcomes, either on an emulated landscape or through microwave synthesis and PXRD.The paper frames practical campaigns around early promising signs, steady improvement, and interpretable suggestions.Source: Introduction · p. 2
  • ActMOF benchmarks Gaussian-process, statistical-ensemble, LLM-only, and hybrid GP-LLM policies on an experiment-grounded landscape, then transfers selected strategies to six related aluminum MOFs.The landscape expands 95 measured microwave syntheses into 6,101,172 encoded conditions.Source: Introduction · p. 2

Results and Discussion

ActMOF benchmarks active-learning policies on an experiment-grounded aluminum-MOF synthesis landscape and tests whether the leading search strategy transfers to related MOFs. Matérn Gaussian processes were among the strongest policies, and benchmark-guided campaigns found crystalline products across six frameworks.

Results and Discussion

  • ActMOF assigned q as diagnostic PXRD peak intensity divided by full width at half-maximum, a first-stage crystallinity ranking rather than downstream stability or sorption validation.The model MOF-321 reaction space encoded precursor and modulator amounts, added solvent, reaction time, and temperature.Source: Results and Discussion · p. 3
  • The benchmark expanded 95 measured outcomes into 6,101,172 candidate conditions using a rule-guided emulator intended to preserve landscape structure, not predict every unmeasured outcome.Because the emulator uses random forests, tree policies might find the landscape unusually easy, but they showed no measurable advantage over GPs.Source: Results and Discussion · pp. 3–4
  • Top-1% q began at 42,463 and top-0.1% q exceeded 100,288; high-q regions favored low metal and modulator, long heating, and elevated temperatures.Added water mattered less significantly, and the favorable parameter window narrowed at stricter q tiers.Source: Results and Discussion · p. 4, Figure 2 · p. 5
  • With repeat-loading noise, PI lost 8.8% of final best q and EI lost 0.3%, so EI was chosen for experimental transfer.The authors attribute the difference to EI rewarding expected improvement size while PI counts any improvement equally.Source: Results and Discussion · p. 5
  • Matérn GPs with PI or EI were among the most reliable policies; tree ensembles were complementary, while GP–LLM hybrids did not overtake GPs and LLM-only proposals lagged.The benchmark comparison plots cumulative best q over 120 experiments; the tree-ensemble comparison also reports top-1% hit rates.Source: Results and Discussion · pp. 5–6, Figure 3 · p. 6
  • Transfer used MOF-321 measurements as a calibrated prior, then adapted a residual GP to each candidate framework’s PXRD responses as new experiments accumulated.The six targets preserved aluminum-based reticular chemistry while varying linker solubility, hydrophilicity, and angularity.Source: Results and Discussion · pp. 7–8
  • Across six 12-run campaigns, every final PXRD pattern matched its simulated target phase; average best-q gain was 2.2-fold, although CAU-23 did not exceed its already high starting result.All 18 prior-informed starting experiments crystallized, as did 67 of 72 experiments overall.Source: Results and Discussion · p. 8, Figure 5 · p. 9
  • Best recipes retained low metal and base amounts and 50–60-minute holds, but temperature optima varied by linker rather than simply copying MOF-321’s optimum.CAU-10 and MTV-MIL-160 converged at 140 °C, while MIL-160, CAU-23, MTV-MOF-303, and MIL-53-muc peaked at 100–110 °C.Source: Results and Discussion · p. 8

Concluding Remarks

The conclusion presents ActMOF as a reproducible, crystallinity-guided benchmark for studying how algorithms navigate sparse reticular-synthesis spaces. It highlights early confidence, controlled Matérn-GP exploitation, and inspectable laboratory prescriptions, with adaptive search principles tested across six aluminum MOFs.

Concluding Remarks

  • ActMOF advances the shift from isolated trial-and-error toward quantitative, interpretable conversion of crystallization outcomes into transferable synthetic knowledge.The benchmark is described as a reproducible platform for studying optimization in sparse-success reticular synthesis spaces.Source: Concluding Remarks · pp. 9–10
  • ActMOF identifies early confidence, Matérn-GP expected-improvement, and inspectable laboratory prescriptions as components of efficient MOF discovery.The conclusion frames these as three features of the approach, rather than a fixed synthesis recipe.Source: Concluding Remarks · p. 10
  • The transferable element is an adaptive prior that accumulates framework-specific evidence, as suggested by validation across six reticularly related aluminum MOFs.The six frameworks named are CAU-10, CAU-23, MIL-160, MIL-53-muc, MTV-MIL-160, and MTV-MOF-303.Source: Concluding Remarks · p. 10

Associated Content

The associated content lists the available ActMOF dataset, benchmark code, interactive synthesis-condition demonstration, and Supporting Information.

Associated Content

  • The ActMOF dataset and benchmark code are available, alongside an interactive demonstration for recommending synthesis conditions.The Supporting Information covers materials and methods, benchmark construction and validation, benchmarking protocols, transfer learning, synthesis details, PXRD patterns, and N2 sorption data.Source: Associated Content · p. 10

Author Information

The author information identifies Zhiling Zheng as the corresponding author and lists the paper's authors, affiliations, and conflict-of-interest statement.

Corresponding Author

  • Zhiling Zheng is listed as the corresponding author, affiliated with the Department of Chemistry and Institute of Materials Science and Engineering.The supplied author information does not include contact details.Source: Corresponding Author · p. 10

Authors

  • The listed authors are Hongrui Sheng, Shutian Liang, Reid A. Coyle, Shyam Chand Pal, Peter Walther, Bin Feng, Xinyu Ding, and Zhiling Zheng.Their affiliations include Washington University, Fudan University, and Imperial College London.Source: Authors · p. 10

Conflicts of interest

Supporting Information

The Supporting Information accompanies ActMOF and identifies its authors and institutional affiliations.

Supporting Information

  • The Supporting Information is for ActMOF: Mapping Reticular Synthesis through Active Learning.Source: Supporting Information · p. 19
  • The listed Supporting Information authors are Hongrui Sheng, Shutian Liang, Reid A. Coyle, Shyam Chand Pal, Peter Walther, Bin Feng, Xinyu Ding, and Zhiling Zheng.Source: Supporting Information · p. 19
  • The authors' listed affiliations span Washington University, Fudan University, and Imperial College London.Source: Supporting Information · p. 19

S2 The ActMOF Landscape

The supporting information organizes ActMOF around its landscape, search strategies, and benchmark results. Its contents list the topics covered within each of these sections.

S2 The ActMOF Landscape

  • The ActMOF Landscape section covers the crystallinity index, anchor experiments, variable space, emulator, reliability, structure, and validation optima.Source: S2 The ActMOF Landscape · p. 19

S3 Search Strategies

  • The Search Strategies section covers the protocol, Gaussian processes, acquisition functions, LLM policies, hybrid GP–LLM policies, and all strategies.Source: S3 Search Strategies · p. 19

S4 Benchmark Results

  • The Benchmark Results section lists metrics, all-strategy and leading-group results, robustness, PI versus EI, warm starts, and no-product-region modeling.Source: S4 Benchmark Results · p. 19

S5 Transfer Learning

The supporting information outlines four components of the transfer-learning section: a desktop application, GP approaches, calibration, and six campaigns.

S5.1 Desktop Application

S5.2 Reference and Residual GP

S5.3 Calibration

S5.4 Six Campaigns

Section S1. Materials and Methods

The methods describe the chemicals and microwave synthesis and workup used in the study, including the validation chemicals and molar masses listed in Table S1.

Materials

Microwave reactor

  • Syntheses used an Anton Paar Monowave 400 microwave reactor at 850 W and 600 rpm; solids were isolated, washed, and measured by PXRD.Samples for N2 sorption were activated.Source: Section S1. Materials and Methods · p. 21

Section S2. The ActMOF Landscape

The section uses measured MOF-321 syntheses to define an emulated crystallinity landscape and situates the best-found conditions for six related validation frameworks within it. It compares these conditions and shared starting conditions with MOF-321 anchors and top tiers, noting that the six best-found conditions fall in MOF-321’s top-1% tier.

S2.1. Crystallinity Index q

  • q is the diagnostic PXRD peak intensity divided by its full width at half maximum, and is comparable only within the same phase measured under the same protocol.Repeated loading gave q a 30.7% coefficient of variation, while crystalline anchors spanned three orders of magnitude; within a phase, high q tracked porosity but was not a measure of surface area.Source: S2.1. Crystallinity Index q · pp. 21–23

S2.2. Anchor Experiments

  • The deduplicated MOF-321 anchor set contained 95 unique conditions: 67 crystalline products and 28 conditions without product.The anchors retained their measured outcomes to train and ground the emulator.Source: S2.2. Anchor Experiments · p. 24, Table S4 · p. 24

S2.3. Variable Space

  • Five integer-coded synthesis variables define a shared finite grid of recipes, while the anchors cover only part of that space and concentrate in the productive region.The later recorded experiments cluster at low AlCl3·6H2O and NaOH volumes, longer hold times, and 90–150 °C.Source: S2.3. Variable Space · pp. 26–27

S2.4. Emulator

  • The emulator predicts peak intensity and FWHM with separate random forests, blends predictions with nearby anchors, and applies output thresholds to classify no-product conditions.In cross-validation it recognized all 28 no-product samples; among crystalline samples predicted as products, median FWHM error was 0.06°.Source: S2.4. Emulator · p. 28, Figure S5 · p. 29

S2.5. Reliability

  • All 28 anchors without product were predicted as q = 0, and crystalline anchors had measured–predicted rank correlation ρ = 0.78.The emulator predicted 57% of crystalline anchors within half to twice their measured q, but missed 9 as no product; it never predicted a product where none was measured.Source: S2.5. Reliability · p. 29, Table S7 · p. 30
  • Every top-1% landscape condition lay within 7 grid steps of an anchor, but 93% of the no-product region was flagged as unsupported extrapolation.A map built from the first 60 anchors ranked the 35 later experiments better than the nearest-anchor baseline (Spearman ρ = 0.61 vs. 0.43); its errors were false zeros, not false products.Source: S2.5. Reliability · pp. 31–32, Table S8 · p. 32, Table S9 · p. 32

S2.6. Structure

  • Only 6.3% of the landscape yielded crystalline product, and just 1% reached q ≥42,463.High-q conditions clustered at low AlCl3·6H2O and NaOH, long hold times, and elevated temperatures; a one-step increase in NaOH from 0.7 to 0.8 mL lowered q from 76,141 to 1,379.Source: S2.6. Structure · pp. 32–33, Table S10 · p. 32

S2.7. Validation Optima

  • All six validation optima fell inside MOF-321’s top-1% tier and within 1–2 grid steps of its top-0.1% tier.A random grid condition was a median 16 and 20 steps from the nearest top-1% and top-0.1% conditions, respectively; the three shared starting conditions were already in the top 1%.Source: S2.7. Validation Optima · p. 34
  • Table S11 reports Chebyshev-unit distances from each validation MOF’s best condition and the three shared starts to MOF-321 anchors and top tiers.Source: Table S11 · p. 34
  • Figure S8 maps the six validation MOFs’ best-found conditions and three shared starts across ten MOF-321 projections, showing the top-1% and top-0.1% tiers; CAU-23’s best condition is R00001.The projections show the highest q over the other three variables and are zoomed to the productive part of the grid.Source: Figure S8 · p. 35

Section S3. Search Strategies

Section S3 describes the shared protocol and search strategies used to explore the ActMOF landscape, including Gaussian-process, acquisition-function, and LLM policies, the role of priors and experimental histories, and the surrogate models and reproducibility controls. It also reports query patterns in gpt-5.5 and hybrid runs.

S3.1. Protocol

  • Until a product appears, model-based policies act like random search; this costly phase motivates starting the 12-experiment wet-lab campaigns from a prior rather than random conditions.Because 94% of the landscape gives no product, models cannot distinguish conditions until results differ. In shared-start runs, half of 20 Matérn-5/2 EI runs needed more than 12 experiments to find a first product; across iterations 1–5, this accounted for 57 of 75 EI queries and 24 of 75 PI queries. The benchmark used three starting experiments and then three per iteration for 120 experiments over 39 iterations, with five runs per strategy.Source: S3.1. Protocol · p. 36

S3.2. Gaussian Processes

  • After 120 experiments, Matérn-5/2 GPs learned shortest length scales for AlCl3 · 6 H2O (median 0.09), temperature (0.40), and NaOH (0.48), while DI water and hold time exceeded their full ranges.The learned length scales indicate that metal and base amounts and temperature determine whether a product forms, whereas added DI water has little effect; its length scale reached the fit's upper bound in four of five runs.Source: S3.2. Gaussian Processes · pp. 38–39, Figure S10 · p. 39

S3.3. Acquisition Functions

  • Matérn-5/2 EI’s 585 proposals yielded products in 59%, top-1% results in 39%, and top-0.1% results in 16%, shifting from broad early exploration toward productive regions.PI rewards any improvement, while EI rewards the expected size of improvement and increases with uncertainty; other tested rules include optimistic estimates, Thompson sampling, and greedy predictions.Source: S3.3. Acquisition Functions · pp. 39–40, Figure S11 · p. 40

S3.4. LLM Policies

  • LLM policies propose three encoded conditions from recent experimental history without candidate lists, while hybrid policies share the same history across GP and LLM iterations.Each request includes the 45 most recent experiments and their q values; the program rejects out-of-range or repeated conditions, retries invalid batches, and uses three random untried conditions after three failed attempts.Source: S3.4. LLM Policies · pp. 41–42
  • gpt-5.5 returned valid batches on the first attempt in 191/195 iterations, while gpt-4o and gpt-5-mini used fallback in 28% and 36% of iterations, respectively.Most rejected answers repeated an earlier experiment: 249 of 259 gpt-4o rejections and all 379 gpt-5-mini rejections.Source: S3.4. LLM Policies · p. 45
  • The LLM runs recorded model snapshots, fixed accepted sampling settings and seeds, tested that settings were applied, and logged iterations so requests and benchmark curves could be rebuilt without API calls.The authors note that backend changes within a snapshot could not be excluded and fallback conditions were not drawn with a fixed seed, so reruns should be similar but not necessarily identical.Source: S3.4. LLM Policies · p. 45
  • Among 585 gpt-5.5 queries in runs 0–4, 81% crystallized, 60% reached the top-1% tier and 41% the top-0.1% tier, versus 59%, 39% and 16% for Matérn-5/2 EI.gpt-5.5 concentrated queries in the known productive window, but its best-so-far rose slowly after initial hits because it rarely tested uncertain conditions outside that window.Source: S3.4. LLM Policies · p. 45, Figure S12 · p. 46

S3.5. Hybrid GP–LLM Policies

  • In the EI hybrid, 42 of 55 best-so-far improvements came from LLM iterations; in the PI hybrid, 38 of 58 did, and the final best came from an LLM iteration in all five runs of both hybrids.The policies alternated GP and gpt-5.5 batches using a shared experiment history, with the GP mainly adding smaller refinements.Source: S3.5. Hybrid GP–LLM Policies · pp. 46–47

S3.6. All Strategies

  • The benchmark included 39 strategies, with 16 representatives compared in the main text; non-GP surrogates included random forests, ExtraTrees, conformal random forests and linear bandits.Tree models supplied uncertainty from prediction spread, with conformal random forests widening it using out-of-bag error; the authors note linear models cannot represent a narrow, closed high-q region.Source: S3.6. All Strategies · p. 47

Section S4. Benchmark Results

Across 39 policies, full GPs, tree ensembles and hybrids generally outperformed random search, while six leading strategies could not be distinguished statistically. Their advantage over random search persisted across rebuilt landscapes, but rankings among the six varied; warm starts improved early progress, whereas explicit modeling of no-product outcomes did not outperform the baseline.

S4.2. All Strategies, n = 5

S4.3. Leading Group, n = 20

  • 85–100% of runs for each of six leading strategies reached the top-0.1% tier after 120 experiments, versus 10% for random search, with no significant differences among the six.After 12 experiments, their early top-1% hit rate was 0.05–0.35, compared with 0.10 for random search.Source: S4.3. Leading Group, n = 20 · p. 55, Figure S16 · p. 55, Table S16 · p. 55

S4.4. Robustness to the Construction of the Landscape

  • Across all 41 rebuilt landscapes, each of the six strategies beat random search, with mean best q at least 1.4 times as high, but each strategy ranked first on at least one landscape.The top-0.1% tier shifted within the productive region, and 75–100% of original top-0.1% conditions remained in the top 1% of each rebuilt landscape.Source: S4.4. Robustness to the Construction… · p. 56, Figure S17 · p. 57, Figure S18 · p. 58

S4.5. PI versus EI

  • PI’s early lead over EI in independent-start runs came from its starting batches: 7 of 20 PI starts contained a product, versus 1 of 20 EI starts; with shared starts, their curves overlapped.With 30% measurement noise, final mean best q fell 8.8% for PI and 0.3% for EI.Source: S4.5. PI versus EI · pp. 58–59, Figure S19 · p. 58, Figure S20 · p. 59

S4.6. Warm Start

  • Warm starts placed all runs in the top-1% tier after 12 experiments, versus at most 35% of cold-start runs; the six strategies reached the top-0.1% tier in all runs by 120 experiments.The prior start and random starts in the top-1%–0.1% band produced almost the same curves; the start located the productive region while the strategy advanced toward the top-0.1% tier.Source: S4.6. Warm Start · pp. 59–60, Figure S22 · p. 61, Table S17 · p. 61

S4.7. Modeling the No-Product Region

Section S5. Transfer Learning

The transfer workflow uses MOF-321 measurements as a calibrated prior, then adapts that prior with new-MOF data and a residual GP. In six validation campaigns, five frameworks improved their best q within 12 experiments, while the transferred trends supplied a productive starting region rather than a shared optimum.

S5.1. Desktop Application

  • The desktop application proposes three experiments per iteration using Matérn-5/2 EI in all six campaigns, with a calibrated prior and its default schedule.Its GP uses a single length scale across five variables because campaigns contain only 3–12 new-MOF results.Source: S5.1. Desktop Application · pp. 63–64, S5.3. Calibration · p. 65

S5.2. Reference and Residual GP

Calibration.

  • Calibration fits an offset and slope to map MOF-321 predictions onto new-MOF responses, penalizing slopes far from 1 while data are sparse.The offset absorbs intensity-scale differences; the slope controls transfer strength, and is restricted to [−2, 3].Source: Calibration · p. 64

S5.3. Calibration

  • For CAU-10 and CAU-23, calibration temporarily inverted the prior after anti-correlated results, then moved back toward positive slopes; the next CAU-10 batch contained its best sample.The correction took one batch, at the cost of a poor batch.Source: S5.3. Calibration · p. 65

S5.4. Six Campaigns

  • Five of six validation MOFs improved their best q 1.23–5.40-fold within 12 experiments; CAU-23 retained its starting condition as its best, and 67 of 72 experiments yielded crystalline product.All best samples except CAU-23’s were found during calibrated-prior rounds 1–3.Source: S5.4. Six Campaigns · p. 65, Figure S24 · p. 66, Table S19 · p. 66
  • The MOF-321 prior transferred general crystallization rules, but each framework’s optimum still required search; five of six best samples followed recalibration.Best recipes were in or near MOF-321’s productive window but differed in optimal time and temperature, and none except CAU-23’s matched a starting condition.Source: S5.4. Six Campaigns · pp. 65, 67

Section S6. Synthesis of the Validation MOFs

The validation MOFs were synthesized using a shared microwave procedure, with linker identity and active-learning-proposed reaction conditions varying across campaigns.

Section S6. Synthesis of the Validation MOFs

  • All six frameworks used the same general synthesis procedure, with the linker as the stated difference.Linkers were dissolved in DI water and NaOH before adding AlCl3 · 6 H2O, followed by microwave heating and workup.Source: Section S6. Synthesis of the… · p. 67
  • Each validation synthesis used 0.1 mmol total linker, or 0.05 mmol of each linker for the two multivariate frameworks, with one encoded-variable step of 10 µL.The validation syntheses were run at one tenth the scale of Table S5.Source: Section S6. Synthesis of the… · p. 67
  • The validation set comprised CAU-10, CAU-23, MIL-53-muc, MIL-160, MTV-MIL-160 and MTV-MOF-303, each prepared from its specified linker or linker pair.The 12 synthesis conditions for each framework are listed in Tables S20–S25.Source: Section S6. Synthesis of the… · pp. 67–68

Section S7. PXRD and Campaign Trajectories

Section S7 documents the 12-synthesis campaigns and their PXRD results across six aluminum MOFs. The tables report recipes and diffraction metrics, while the figures show patterns in synthesis order.

Campaign trajectories

  • Each MOF campaign comprises 12 syntheses in four rounds; the first three conditions came from the MOF-321 reference prior, and later rounds used the calibrated prior with residual GP.The tables report measured intensity, FWHM, and q = I/FWHM; the campaign’s best q and its improvement over the starting experiments are summarized elsewhere.Source: Section S7. PXRD and Campaign… · p. 68

Campaign results

Campaign trajectories

  • One CAU-10 recipe used 170 µL rather than the proposed 190 µL of DI water because of a pipetting error; the condition actually used was listed and supplied to the model.Source: Section S7. PXRD and Campaign… · p. 70

Campaign results

Section S8. Nitrogen Sorption

Nitrogen sorption compared microwave campaign optima with reflux products for CAU-23 and MIL-160. The microwave products were porous, with CAU-23 near its literature BET area and MIL-160 reaching 80% of its reflux BET area.

Measurement and comparison

  • BET and Langmuir areas were compared using six adsorption-branch points over p/p0 = 0.05–0.10, with R2 ≥0.9997 for every sample and equation.The microwave samples were the campaign optima; reflux samples were 6 h products.Source: Section S8. Nitrogen Sorption · p. 75
  • Table S26 reports BET and Langmuir areas, in m2 g−1, from the adsorption branch at p/p0 = 0.05–0.10.Source: Table S26 · p. 75

Results

  • CAU-23’s microwave optimum had SBET = 1173 m2 g−1, 94% of the literature BET value, while the reflux product measured 56 m2 g−1.The reflux product was mistakenly heated overnight and converted into a different, nearly non-porous phase.Source: Section S8. Nitrogen Sorption · p. 75
  • MIL-160’s microwave optimum reached 80% of the 6 h reflux BET area: 729 against 908 m2 g−1.The corresponding Langmuir areas were 850 against 1058 m2 g−1.Source: Section S8. Nitrogen Sorption · p. 76
Loading 2610.19404v1…