The Flow has not summarised this story yet — read it at arXiv Machine Learning.
arXiv:2607. 19114v1 Announce Type: new Abstract: Activities in aqueous electrolyte solutions, usually described by ionic activity and osmotic coefficients, are important properties for modeling many processes in industry and nature.
The paper presents a hierarchical deep‑learning framework that integrates compositional, structural, and transport models to screen solid‑state electrolytes. Four modules—L‑G‑DCNN, DenseGNN, MatterSim, and DeePMD—coordinate to evaluate thermodynamics, multi‑property performance, and kinetic transport, outperforming existing methods. Applied to over 30 million candidates, the workflow identifies 97 high‑performance materials, mainly halides, and links Li⁺ jump‑network connectivity to ionic conductivity while highlighting limits for oxide electrolytes.
The article examines how scientific literature, often used as a data source for AI in materials science, can contain hidden inaccuracies such as text-figure mismatches, ambiguous axis labels, unit inconsistencies, and missing measurement context. By tracing solid electrolyte conductivity values from original papers to curated datasets, the authors uncover recurrent errors that are numerically plausible yet hard to detect, leading to significant label noise in AI models. A cross-database example demonstrates that ambiguous reporting can cause a 100‑fold error in conductivity values, underscoring the need for traceable reporting, rigorous curation, and validation practices in AI-driven discovery.
ProtoMI is a literature‑driven framework that learns structural priors from 126 reported boron‑containing electrolyte additives and applies them to screen 179,977 unlabeled candidates. Using graph contrastive learning, it identifies seven interpretable prototypes and adapts them through semi‑supervised contrastive learning, achieving enrichment factors of 9.2–45.6 while evaluating less than 2% of the candidate space. The method led to the discovery of four commercially accessible additives, including TNDB, which improves high‑temperature LiFePO4||graphite cycling by 34.93% and forms protective interphases that suppress solvent decomposition and Fe deposition.
arXiv:2606. 08148v1 Announce Type: cross Abstract: Interatomic potentials (IPs) enable large-scale atomistic simulations beyond the reach of first-principles methods, but their predictive reliability depends critically on the selection of training data, quantified uncertainty, and model expressiveness.
arXiv:2606. 30961v1 Announce Type: cross Abstract: Advances in deep learning architectures and representations have enabled ML-driven chemical property prediction, but state-of-the-art (SOTA) models have remained largely confined to independent codebases and lack support for diverse chemical species.