1. Knowledge-Guided Autonomous Discovery of Microenvironment-Tuned Metal–Organic Framework Photocatalysts1
A literature-guided language-model workflow proposed and tested 31 modified metal–organic framework photocatalysts across six automated cycles. Its best material improved hydrogen-production activity approximately 36-fold over the parent in the reported comparison. The study connects cross-domain hypotheses to experimental feedback. Reporting is limited to the publisher-deposited abstract; performance in a differently illuminated reactor is a separate comparison.1
2. Chemistry-Informed Multimodal Model for Structure Elucidation in Automated Reaction Discovery2
The unreviewed DynaMAIK preprint combines mass spectra with reactant and reagent descriptions to identify reaction products. It reports 87% top-choice structure accuracy on a held-out experimental set and an application to high-throughput thiolation. Comparisons with single-input models support complementary chemical constraints, addressing an analytical bottleneck in automated discovery. This account is limited to the deposited abstract.2
3. A Systematic Approach to Batch Active Learning for Material Optimization3
An unreviewed study compares batch active-learning strategies across simulated landscapes, models and dimensionalities, then tests its findings on automated coacervate optimisation. A mostly exploitative strategy with a small exploratory share performed competitively, while efficient batch sizes depended on operational constraints. The comparison informs experiment allocation under fixed budgets. Our access was limited to the deposited abstract.3
4. Agentic Workflow for High-Throughput Synthesis Campaigns: Mapping the Persistent Phosphor Ca2ZnSi2O7:Eu2+,Nd3+4
A language-model agent analysed a 48-sample phosphor synthesis campaign using automated diffraction fitting and spectroscopy. The unreviewed study identified a rare-earth-bearing phase missed in earlier manual analysis, but recovering it required human correction of a proxy structure and re-screening. The emitting europium site remains unresolved. The deposited abstract supports a useful, explicitly supervised campaign-analysis workflow.4
5. KRstereo: Predicting β-Hydroxy Stereochemistry in Polyketides Using Protein Language Models5
KRstereo uses protein-language-model embeddings to predict the stereochemistry produced by polyketide ketoreductase domains. The publisher abstract reports improvement over sequence-motif rules and validation on newly characterised domains in a later MIBiG release. This connects biosynthetic sequences to likely product stereochemistry and supports genome mining. Our access was restricted to the deposited abstract; large-scale annotations remain predictions.5
References
Yiming Zhao; Tao Song; Linjiang Chen; Yan Huang; Kang Sun; Wentao Han; Mingyang Shen; Chenwei Mao; Peng Lan; Meng Zhou; Weiwei Shang; Jun Jiang; Hai-Long Jiang. Knowledge-Guided Autonomous Discovery of Microenvironment-Tuned Metal–Organic Framework Photocatalysts. Journal of the American Chemical Society; 2026; Peer-reviewed journal article; first online. DOI: 10.1021/jacs.6c13252. Accessed 2026-09-20T16:23:23.852Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.1021/jacs.6c13252
Designing second-sphere microenvironments that promote proton-coupled electron transfer is central to catalysis yet difficult to achieve in porous solids, such as metal–organic frameworks (MOFs). Here, we report an end-to-end workflow that couples literature-guided large-language-model (LLM) reasoning with real-time experimental feedback to propose, test, and refine microenvironment designs in MOF photocatalysts. The system mined and fused three domains (namely, photocatalytic H2 production, hydrogenases and enzyme-mimetic catalysis) and deduced the hypothesis that placing basic, hydrogen-bonding groups near catalytic centers would facilitate water activation and proton transfer. The hypothesis was instantiated by postsynthetic modification of UiO-67, generating 31 Pt@UiO-67-X variants and evaluating them across six closed-loop iterations on an automated platform. The search converged on Pt@UiO-67-30 (8-quinolinecarboxylic acid), which delivered 2.33 mmol g–1 h–1, a ∼36-fold improvement over the parent material; in a larger, optimally illuminated reactor the same catalyst reached 12.48 mmol g–1 h–1 while preserving the library’s rank order. Photoluminescence quenching, enhanced photocurrent, and reduced impedance are consistent with faster charge separation, and first-principles calculations are consistent with reduced proton-transfer barriers via N···H hydrogen-bond networks. These results establish a practical microenvironment-engineering strategy in MOFs and show how LLM-guided knowledge fusion with experiment-in-the-loop reasoning can systematize and accelerate targeted discovery of functional materials.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API. Full text not inspected.
Maik G. Niedziella; Philipp M. Pflüger; Ajnabiul Hoque; Frank Glorius. Chemistry-Informed Multimodal Model for Structure Elucidation in Automated Reaction Discovery. ChemRxiv; 2026; Preprint v1; not peer reviewed. DOI: 10.26434/chemrxiv.15008991/v1. Accessed 2026-09-20T16:23:23.853Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.26434/chemrxiv.15008991/v1
Rapid and reliable structure elucidation remains a central bottleneck in automated reaction discovery, where highthroughput experimentation can generate large numbers of crude reaction mixtures faster than they can be interpreted. Gas chromatography coupled with mass spectrometry (GC-MS) provides rapid and information-rich analytical data, but identification of unknown reaction products still relies heavily on expert interpretation or spectral library matching, limiting applicability to novel compounds. Here we introduce Dynamic Mass-Aware Identification from Reaction Knowledge (DynaMAIK), a multimodal transformer framework for reaction-aware structure elucidation from GC-MS data. DynaMAIK combines electron ionization mass spectra with reaction context encoded as reactant and reagent SMILES to predict product structures and molecular formulas. Trained on more than three million reaction–spectrum pairs derived from curated reaction databases, spectral simulation and experimental spectra, DynaMAIK achieved 87% top-1 and 94% top-10 structure accuracy on a held-out experimental test set. Comparisons with spectrum-only and reaction-only models, together with ablation and scrambling experiments, show that spectral evidence and reaction-derived context provide complementary constraints for product identification. Application to high-throughput thiolation reactions demonstrates accurate product assignment directly from experimental screening data. DynaMAIK establishes context-aware GC-MS interpretation as a scalable route toward automated analytical feedback in reaction discovery workflows.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API. Full text not inspected.
Andrea Gardin; Willem van den Hout; Yannick H. A. Leurs; Luc Brunsveld; JanC. M. van Hest; Francesca Grisoni. A Systematic Approach to Batch Active Learning for Material Optimization. ChemRxiv; 2026; Preprint v1; not peer reviewed. DOI: 10.26434/chemrxiv.15008883/v1. Accessed 2026-09-20T16:23:23.853Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.26434/chemrxiv.15008883/v1
Closed-loop automated experimentation platforms are accelerating materials characterization and discovery, but their efficiency hinges on a question that automation alone does not answer: given a fixed evaluation budget, which experiment configurations should be run next? Batch active learning addresses this by selecting a group of candidates at each iteration, yet the design choices that govern its efficiency, how many candidates to acquire per cycle, and how to balance exploitation against exploration remain poorly understood. Here, we present a systematic in-silico benchmark of batch active learning across optimization landscapes, prediction models, and dimensionalities, introducing composite metrics that track optimization progress and surrogate accuracy as co-equal outcomes. We find that an exploitative-dominant strategy with a small fixed exploratory minority matches or exceeds purely exploitative configurations while outperforming purely exploratory ones across landscapes, models, and dimensionalities, and that batchand cycle-efficient regimes are distinct and constraint-dependent. We further validate these findings on an automated coacervate condensate optimization platform, confirming their transfer to a real, noisy materials system.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API. Full text not inspected.
Ziyang Jiang; Eric Riesel; Ian Naccarella; Taylor D. Sparks. Agentic Workflow for High-Throughput Synthesis Campaigns: Mapping the Persistent Phosphor Ca2ZnSi2O7:Eu2+,Nd3+. ChemRxiv; 2026; Preprint v1; not peer reviewed. DOI: 10.26434/chemrxiv.15009032/v1. Accessed 2026-09-20T16:23:23.853Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.26434/chemrxiv.15009032/v1
The structural and spectroscopic record of a high-throughput synthesis campaign exceeds what experimenters can closely and accurately analyze with conventional tools alone. We report a workflow built for this problem: a language model agent operates an automated Rietveld phasesearch pipeline, Crystalyze Match, built on the open-source DARA engine, and then interprets the phases, optical spectra, and synthesis metadata. We demonstrate this workflow on a 48 sample campaign for the persistent phosphor Ca 2 ZnSi 2 O 7 :Eu 2+ ,Nd 3+ (doped hardystonite), a water-stable candidate alternative to SrAl 2 O 4 :Eu 2+ ,Dy 3+ for luminescent road markings. Within the campaign’s lifetime, the workflow mapped two decomposition pathways, incomplete reaction and zinc volatilization, that validate the two-step synthetic approach. The workflow also identified the fate of the rare-earth dopants: a silicate oxyapatite, Ca 2 (Eu,Nd) 8 (SiO 4 ) 6 O 2 , missed by prior manual analysis, is present in every doped sample. Human adjudication proved necessary as the search first scored the oxyapatite through a lanthanide-free proxy structure that scatters too weakly to register at low fractions, and a corrected-model re-screen of all 48 patterns recovered it. Though the rare-earth dopants appear to end up in the oxyapatite phase, emission rises monotonically with Eu content, and the material shows the target persistent luminescence, so where the emitting Eu 2+ resides remains unresolved. Our work highlights that human review and re-screening after every structure-model revision dramatically improve the trustworthiness of campaign-scale claims.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API. Full text not inspected.
Hsin-Ying Tsai; Wenqiang Xu; Wen Jun Xie; Yousong Ding. KRstereo: Predicting β-Hydroxy Stereochemistry in Polyketides Using Protein Language Models. Journal of Chemical Information and Modeling; 2026; Peer-reviewed journal article; first online. DOI: 10.1021/acs.jcim.6c02706. Accessed 2026-09-20T16:23:23.853Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.1021/acs.jcim.6c02706
Polyketides are a major class of bioactive natural products whose activities are often determined by the stereochemistry of β-hydroxyl groups. In type I polyketide synthases (PKSs), ketoreductase (KR) domains establish these stereocenters, making accurate prediction of KR stereochemistry important for natural product discovery and PKS engineering. Existing rule-based methods rely on a limited set of sequence motifs and often perform poorly across phylogenetically diverse taxa. Here, we present KRstereo, a machine learning framework that predicts KR stereochemistry directly from sequence. Analysis of β-modular KR domains from MIBiG 3.1 identified informative sequence features beyond canonical motifs, motivating the use of protein language model embeddings. KRstereo achieved accuracies of up to 95.7% across taxa and 93.0% for non-Streptomyces KRs, consistently outperforming existing rule-based approaches. Validation using newly characterized KR domains from MIBiG 4.0 confirmed strong generalizability, including cases misclassified by current methods. Application of KRstereo to 20,840 β-modular KR domains from antiSMASH enabled large-scale stereochemical annotation of previously uncharacterized PKS systems. By linking sequence to stereochemical function, KRstereo improves reconstruction of polyketide structures from biosynthetic gene clusters and facilitates stereochemistry-aware genome mining and engineering of PKS assembly lines.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API. Full text not inspected.
Publication record
Published 2026-09-20.
Sources, selection and claims were checked in an independent AI editorial review, followed by the AI editor's approval. This is not academic peer review.
- 2026-09-20 · Version b8247375 · Viewing this version