Cancers doi: 10.20517/aiagent.2025.13

Authors: Giyeok Lee,Catherine Stampfl,Aloysius Soon

In computational materials science, global optimization is pivotal for bridging theory and experiment but can fail when the theoretical treatment defining the potential energy surface does not accurately predict stability trends. Conventional approaches to address this rely on statistical sampling over numerous independent calculations or the use of more expensive theories throughout the global optimization process, both of which substantially increase computational cost. To overcome this, we present nature-inspired algorithm for robust atomic structure search (NARA), a framework that combines a firefly algorithm-based multimodal search with uncertainty-aware active learning. Instead of converging to a single structure, NARA simultaneously explores multiple distinct configurations, thereby mitigating sensitivity to potential limitations. For the “8” surface oxide on Cu(111), it achieves higher efficiency than the widely used basin-hopping algorithm. For gold clusters, a single run recovers both planar and non-planar structures, resolving stability reversals induced by different theoretical treatments. NARA thus achieves both efficiency and robustness for reliable atomic-structure identification.

Cancers doi: 10.20517/aiagent.2025.12

Authors: Qingrong Liu,Hao Shen,Hua Meng,Fanyue Qian,Yuting Yao,Yuan Gao,Tingting Xu,Yingjun Ruan

With the integration of renewable energy and energy storage in integrated energy systems, their operational and managerial complexity has substantially escalated. This study introduces a novel operational optimization strategy model, convolutional neural network (CNN)-multi agent twin delayed deep deterministic policy gradient (MTD3), based on deep reinforcement learning (DRL). By integrating expert knowledge into DRL, the challenge of failing to shut down certain equipment, which arises when DRL is applied to control continuous actions, has been addressed. Additionally, it mitigates the inappropriate exploration of agents in dynamic load scenarios. The k-means method is used to categorize the annual load, and train specific agents to handle the classified loads. Additionally, a CNN is proposed for load classification and agent selection. Expert knowledge constraints are incorporated into the reward functions. The CNN-MTD3 method not only improves training speed but also reduces annualized operating costs by 4.7% and 10.4% under cooling and heating load scenarios, respectively, compared to the baseline TD3 (twin-delayed deep deterministic policy gradient) method. Notably, the regulation of battery and thermal energy storage equipment by CNN-MTD3 is particularly significant. In continuous day cooling and heating scenarios, the effective operating h of the battery energy storage system increased by 26% and 98%, respectively. Furthermore, there was a 269.2% increase in thermal energy storage system operating in heating scenarios. We conducted a sensitivity analysis on the number of clusters and the CNN classification within CNN-MTD3 to verify the robustness of the method. These outcomes compellingly underscore the efficacy of the methodology proposed in this study.

Cancers doi: 10.20517/aiagent.2026.03

Authors: Xue Jia,Di Zhang,Yiming Lu,Qian Wang,Hao Li

We introduce StableOx-Cat, an artificial intelligence (AI)-agent framework that enables systematic and reliable exploration of stable metal oxide (MO) electrocatalysts via a unified natural-language interface. StableOx-Cat integrates a large language model (LLM) for intent understanding and task orchestration with deterministic, physics-based analysis tools for electrocatalysis evaluation. User queries expressed in natural language are automatically parsed into structured actions, including database statistics, bulk thermodynamic stability screening based on energy-above-hull criteria, and aqueous electrochemical stability analysis under user-defined pH values and electrochemical potential windows. By applying the physical criteria to screen the stable MO electrocatalysts, StableOx-Cat avoids hallucinations and ensures a physically based stability analysis. This Agent enables the assessment of aqueous electrochemical stability across a wide range of reactions, with applied potentials spanning -2 to 2 V versus standard hydrogen electrode and pH values ranging from 0 to 14. Representative use cases demonstrate how StableOx-Cat enables flexible stability screening of MOs under both thermodynamic and aqueous environments. In addition, the agent architecture supports integration with different LLMs for task execution and query parsing. Overall, StableOx-Cat provides an accessible platform for stability-oriented materials exploration, offering a practical pathway to accelerate the discovery of experimentally relevant MO electrocatalysts for electrochemical applications, and can be generalized to other classes of electrocatalysts, such as alloys, metal nitrides, and carbides.

Cancers doi: 10.20517/aiagent.2026.04

Authors: Yuyang Hong,Xin Mao

A major bottleneck in artificial intelligence (AI)-driven materials discovery is not model architecture, but limited data accessibility: critical experimental knowledge remains locked in figures, heterogeneous reporting formats, and unstructured PDFs. A recent study by Li et al. addresses this challenge by introducing DIVE (Descriptive Interpretation of Visual Expression), a multi-agent extraction framework that transforms figure-centric scientific content into structured, machine-actionable data. Applied to solid-state hydrogen storage materials, DIVE demonstrates substantial extraction gains over conventional direct large language model (LLM) parsing, then scales to mine 4,053 publications (1972-2025) and build a > 30,000-entry database that powers a downstream inverse-design agent, DigHyd. This work offers a practical blueprint for moving from “LLM-assisted reading” to “AI-enabled discovery infrastructure”, linking literature mining, quality scoring, database construction, and target-driven candidate generation in a single workflow.