1. Efficient Free-Energy Landscape Construction Using Graph-Based Active Path Sampling1
An unreviewed preprint connects Gaussian-process active learning with graph paths to reconstruct free-energy landscapes while keeping sampled regions connected. Tests covered analytical landscapes and polymer-grafted nanoparticles in a polymer melt. The approach addresses a practical sampling constraint; its reported efficiency measure is connected-path length. Our assessment is limited to the repository-deposited abstract, rather than an independent runtime benchmark.1
2. Public Database Leakage Distorts Model Rankings in Real-Spectra NMR Structure Elucidation2
A leakage-controlled NMR benchmark found that removing exact training-corpus matches substantially changes the interpretation of structure-prediction performance on real spectra. Molecular-formula and carbon-peak consistency checks then improved ranking. The study makes dataset identity and decoding checks part of the experimental-generalisation question. Reporting is limited to the publisher abstract; the reported result concerns the tested released models.2
3. Artificial Intelligence-Assisted Dopant Discovery toward Air-Stable Sulfide Solid-State Electrolytes3
An unreviewed study combines language-model retrieval with composition-based prediction to screen sulfide-electrolyte dopants. A bismuth-fluoride formulation retained 91% of its ionic conductivity after six hours at 10% relative humidity, with subsequent cell tests also reported. That explicitly bounded exposure is useful evidence for air-stability design. Our access was limited to the repository-deposited abstract.3
4. Uni-Macro-FRPN: Full-Resolution and Cross-Scale Learning for Polymers4
The unreviewed Uni-Macro-FRPN preprint represents both monomer chemistry and complete polymer-chain topology using BigSMILES. It outperformed tested baselines on block-copolymer classification and a topology-rich simulation benchmark, while monomer-focused models remained competitive for linear homopolymers. The comparison identifies when chain organisation adds information. This account is limited to the repository-deposited abstract and preserves the simulated benchmark distinction.4
5. The Interference Index: Quantifying and Improving Error Cancellation in Machine-Learned Thermodynamic Stability Predictions5
An unreviewed preprint separates accurate formation energies from accurate derived stability predictions. Its interference index quantifies error cancellation, and a reaction-aware training objective improved decomposition-energy prediction and stability classification on Materials Project data. The work targets a consequential gap between fitting compounds and comparing reactions. Reporting is limited to the repository-deposited abstract, with no claim of experimental validation.5
References
Mohsen Farshad; Gaurav Arya. Efficient Free-Energy Landscape Construction Using Graph-Based Active Path Sampling. ChemRxiv; 2026; Preprint v1; not peer reviewed. DOI: 10.26434/chemrxiv.15009038/v1. Accessed 2026-09-20T16:15:23.328Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.26434/chemrxiv.15009038/v1
Constructing free-energy landscapes without prior knowledge of their structure often requires dense sampling of the configuration space, leading to prohibitive computational costs. Here, we introduce an efficient graph-based active-learning framework that adaptively constructs free-energy landscapes while preserving the connectivity required for free-energy determination. The method employs a Gaussian-process surrogate to identify informative configurations and connects each newly sampled set of points to the existing sampled region through a shortest-path search on a graph, enabling efficient, connected exploration required for accurate free-energy determination. We validate the framework using analytical two-and three-dimensional model energy landscapes as well as a many-body free-energy landscape derived from molecular dynamics simulations of polymer-grafted nanoparticles in a polymer melt. Across all cases, the method accurately recovers the multidimensional landscape at much lower effective sampling cost than dense or predetermined grid-based sampling, with sampling efficiency quantified by the cumulative length of the connected sampling tree constructed during active learning. More broadly, this work establishes a general framework for data-efficient exploration of high-dimensional free-energy landscapes in molecular and materials simulations.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API; full text not inspected.
Zihan Zhang; Mengxi Chen; Xuezhou Zhao; Yutao Guo; Dan Wu. Public Database Leakage Distorts Model Rankings in Real-Spectra NMR Structure Elucidation. Journal of Chemical Information and Modeling; 2026; Peer-reviewed journal article; first online. DOI: 10.1021/acs.jcim.6c02552. Accessed 2026-09-20T16:15:23.328Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.1021/acs.jcim.6c02552
Abstract Sequence models that translate NMR spectra into molecular structures report up to 96% top-1 accuracy, but they are pretrained on simulated spectra and tested on public databases that can overlap those corpora. We build a leakage-controlled benchmark on nmrshiftdb2 and find that 26.9% of public molecules are exact training-corpus matches under a stated identity rule, with substantial training-set proximity remaining after exact-match removal. On the wider exact-match-removed real-spectrum cohort, released-model top-1 accuracy is 24.4% (95% CI 23.3–25.6%). A training-free rerank using molecular formula and 13C-peak count consistency raises it to 31.1% (29.9–32.4%), supporting leakage-controlled evaluation and decode-time consistency checks before claims of experimental generalization for simulation-pretrained inverse models.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API; full text not inspected.
Zhuo Chen; Lin Hong; Xi Zhang. Artificial Intelligence-Assisted Dopant Discovery toward Air-Stable Sulfide Solid-State Electrolytes. ChemRxiv; 2026; Preprint v1; not peer reviewed. DOI: 10.26434/chemrxiv.15009059/v1. Accessed 2026-09-20T16:15:23.328Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.26434/chemrxiv.15009059/v1
Sulfide solid-state electrolytes (SSEs) are highly promising candidates for all-solid-state batteries (ASSBs), but the severe degradation triggered by moisture exposure remains a major obstacle to their large-scale application. Despite compositional doping could effectively enhance the air stability, there is currently no relevant screening strategy to support the efficient exploration of dopants. Herein, we develop an artificial intelligence-assisted dopant screening platform for air-stable sulfide SSEs (LPSC). Retrieval-augmented system combined with large language model could effectively prescreen the dopants that enhance both the ionic conductivity, and interfacial compatibility of the electrolyte. More importantly, the impact of dopants on the air stability of electrolytes was quantified by the minimum Gibbs free energy change () of the hydrolysis reaction, which could be predicted by a composition-based machine learning model. Guided by this strategy, several promising dopants, including BiF3, CoF3, InF3, and SnF4 were successfully identified. As a proof of concept, LPSC-BiF3 electrolyte demonstrates superior air stability, delivering a high ionic conductivity retention of 91% after exposure to air with 10% RH for 6 h. Consequently, the exposed LPSC-BiF3 electrolyte enabled excellent cycling stability in Li-In||NCM811 full cells. The proposed screening strategy significantly promotes the exploration of high-performance dopants for air-stable sulfide SSEs.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API; full text not inspected.
Jintao Wu; Yiran Shan; Rui Zhang. Uni-Macro-FRPN: Full-Resolution and Cross-Scale Learning for Polymers. ChemRxiv; 2026; Preprint v1; not peer reviewed. DOI: 10.26434/chemrxiv.15009049/v1. Accessed 2026-09-20T16:15:23.328Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.26434/chemrxiv.15009049/v1
Polymer properties are governed by interactions across scales. Existing polymer models commonly retain either monomer chemistry without the polymer graph, or the polymer graph with simplified monomer chemistry. Here, we present Uni-Macro-FRPN (FRPN), a Full-Resolution Polymer Network unifying atom-level and within-monomer structural encoding with an explicit monomer-instance graph of a polymer chain. Its two Transformers jointly learn atom-informed monomer semantics, sequence order, and chain topology. Using BigSMILES to construct monomer semantics and polymer graphs, FRPN achieves 86.4% accuracy and 90.6% ROC–AUC on Block Copolymer Database (BCDB) lamellar-versus-non-lamellar classification, outperforming the evaluated baselines and supporting joint reasoning over monomer semantics and chain structure. Ablation results support a contribution from joint representation beyond the tested increases in parameter count. On the linear homopolymer benchmark, monomer-centric learning remains competitive, identifying a boundary case in which polymer-scale organization is relatively simple. In addition, existing models are primarily developed around linear polymers. To test whether this full-chain representation generalizes beyond linear polymers, we construct an all-atom moleculardynamics (MD) benchmark of 1640 datapoints spanning diverse monomer chemistries, sequence orderings, chain topologies, and physical properties. FRPN achieves the strongest overall performance on this topology-rich benchmark, and the accompanying diagnostics suggest a benefit from jointly modeling monomer chemistry and polymer structure within one architecture. Taken together, FRPN provides a practical route for moving polymer representation learning beyond SMILESbased descriptions by using BigSMILES to connect monomer semantics with explicit polymer-chain topology, and establishes a foundation for full-chain, cross-scale polymer modeling. Its leading overall performance among the evaluated models on both the real-world BCDB benchmark and the topology-rich MD benchmark further suggests a promising direction for polymer informatics: future polymer prediction models should treat polymers not only as collections of monomer descriptors, but as complete multiscale chemical and topological objects.
Abstract only: original publisher/repository-deposited metadata retrieved from official Crossref API; full text not inspected.
Alexander Kanzow; Cesare Boriosi; Carlos R. Jacinto-Mejía; Loriano Storchi; Giovanni Bistoni. The Interference Index: Quantifying and Improving Error Cancellation in Machine-Learned Thermodynamic Stability Predictions. ChemRxiv; 2026; Preprint v1; not peer reviewed. DOI: 10.26434/chemrxiv.15008797/v1. Accessed 2026-09-20T16:15:23.328Z.
Source evidence and access
Publisher-deposited abstract, https://api.crossref.org/works/10.26434/chemrxiv.15008797/v1
Machine-learned formation enthalpies for inorganic materials approach density functional theory (DFT) accuracy, yet thermodynamic stability predictions derived from them can remain unreliable, a gap commonly associated with error cancellation in DFT that independent per-compound training does not reproduce. We introduce the interference index ξ, a dimensionless measure of how strongly errors reinforce or cancel in linearly derived quantities. Its root-mean-square average, ξrms, provides a reaction-size-independent null reference, with ξrms = 1 for the isotropic null model. Across seven published models benchmarked on Materials Project compounds, error-cancellation behaviour is distinct from per-compound accuracy and becomes progressively less favourable with increasing formation-enthalpy accuracy among the compositional models. We then introduce iiLoss, a reaction-aware objective that favours cancellation of formation-enthalpy errors rather than directly minimising decomposition-enthalpy error. On Materials Project data, iiLoss reduces decomposition-enthalpy error by 21% and improves thermodynamic stability classification by 6.8% in F1 score, outperforming direct supervision on decomposition enthalpies. These results show that error cancellation can be quantified and directly optimised to improve derived thermodynamic predictions.
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 1b445815 · Viewing this version