Ni-rich cathode materials are currently widely used in electric vehicles and energy storage systems due to their high energy density. However， as nickel content increases， cycle stability deteriorates significantly. First， high-nickel materials inherently suffer from cation misalignment during high-temperature sintering and electrochemical cycling， which readily leads to the formation of non-electrochemically active rock salt phase structures. Furthermore， the extraction and insertion of lithium ions during charge-discharge cycles are often accompanied by anisotropic volume changes triggered by the H₂-H₃ phase transition， inducing particle cracking and resulting in intergranular and intragranular cracks. Subsequently， the exposure of fresh surfaces due to these cracks intensifies side reactions with the electrolyte. Particularly at high voltages of 4.6 V， deep lithium-ion deintercalation generates a large amount of highly oxidizing Ni⁴⁺ on the material surface， triggering the release of lattice oxygen. This also induces severe surface structural reorganization， leading to the formation of a thick cathode-electrolyte interphase （CEI） layer， which impedes the interfacial transport kinetics of Li⁺ and ultimately results in the mechanical and electrochemical failure of the material. To address the issue of performance degradation under high-voltage cycling， LiNi0.89Co0.05Mn0.05Al0.01O2 （NCMA-ZrF） material was prepared via zirconium-fluorine co-doping. This approach effectively mitigated capacity loss caused by structural degradation during long-term cycling and significantly improved electrochemical stability under harsh conditions such as high voltage. Zr and F elements were introduced into the material via high-temperature solid-state doping to modify its properties. The strong Zr—O bond energy effectively stabilizes the crystal lattice framework. Furthermore， the more electronegative F was utilized to synergistically reinforce the oxygen framework constructed by Zr， thereby reducing the reactivity of lattice oxygen and suppressing particle cracking and interfacial side reactions during high-voltage cycling. Consequently， the structural and interfacial stability of the material were enhanced， thereby suppressing irreversible phase transformations caused by oxygen evolution under high voltage at the source. Under testing conditions of 4.3 V， 1 C， and 30 °C， NCMA-ZrF exhibited a capacity retention rate as high as 99.97% after 100 cycles. To achieve higher energy density， the cut-off voltage was increased to 4.6 V. NCMA-ZrF still demonstrated excellent specific discharge capacity and good structural stability， with a first-cycle specific discharge capacity of 236.91 mAh/g at 0.1 C. Even after 100 cycles at 1 C， the capacity retention rate reached 81.91%. Furthermore， at 4.3 V， NCMA-ZrF demonstrated good practical performance under high-rate and high-temperature conditions： at 5 C， its specific discharge capacity reached 181.15 mAh/g， with a capacity retention rate of 95.11% after 200 cycles； even after 50 cycles at 45 °C， the capacity retention rate remained as high as 99.18%. This study utilized Zr/F co-doping to effectively enhance the structural stability of high-nickel materials， providing a new technical route for the development of next-generation high-performance lithium-ion batteries.
The integration of Large Language Models （LLMs） with agent architectures is reshaping educational technology by transforming language-generation systems into interactive agents that can perceive learner states， organize domain knowledge， conduct multi-turn pedagogical dialogue， select instructional strategies， and evaluate learning outcomes.This survey reviews recent research on LLM-based educational agents and clarifies their application scope， technical architecture， evaluation methods， and unresolved challenges.Following a systematic literature review and inductive analysis， we collected studies mainly from Web of Science， Scopus， IEEE Xplore， ACM Digital Library， Google Scholar， arXiv， and CNKI， with emphasis on work published from 2022 to 2026， when generative artificial intelligence and agent-based educational applications developed rapidly.At the application level， the reviewed systems are grouped by service object into student-oriented， teacher-oriented， and classroom-oriented agents.Student-oriented agents cover personalized tutoring， automated assignment scoring， and learning diagnosis； representative systems show progress in Socratic tutoring， path planning， rubric-based assessment， explanatory feedback， knowledge tracing， and knowledge-gap identification.Teacher-oriented agents focus on teacher training and lesson preparation by simulating student roles， generating teaching materials， producing structured questions， and supporting quality review.Classroom-oriented agents extend LLM support from one-to-one tutoring to multi-role learning environments by simulating teachers， teaching assistants， and peer learners in classroom or MOOC-like settings.The reviewed empirical evidence reports several notable results： multi-agent scoring can improve agreement with human raters， with Quadratic Weighted Kappa gains of 34.86% in one representative framework； engineering feedback generation has been tested on a dataset of about 5800 samples， with the best model reaching 58.7% accuracy on unseen questions； and multimodal affect detection systems such as MITERS report emotion-recognition accuracy of 97% under experimental conditions.At the technical level， the paper abstracts educational-agent implementation into five layers： perception， knowledge， dialogue， pedagogical strategy， and learning-outcome evaluation.These layers correspond respectively to affective and cognitive state sensing， knowledge graph or retrieval-augmented knowledge organization， instructionally aligned dialogue， adaptive intervention planning， and automated or human-centered assessment.At the evaluation level， the survey summarizes three complementary paradigms： programmatic evaluation， human judgment， and LLM-as-a-judge evaluation， while emphasizing that educational validity cannot be reduced to generic natural-language-generation metrics.Overall， LLM-based educational agents have made substantial progress in functionality， scalability， and interaction richness， but their deployment in real educational settings is still constrained by hallucination control， weak long-term evidence， cold-start diagnosis， insufficient robustness in authentic classrooms， and the lack of standardized evaluation frameworks.Future research should strengthen longitudinal studies， integrate learning-science theory with agent design， improve transparent and controllable knowledge grounding， and establish evaluation protocols that jointly measure accuracy， pedagogical quality， learner agency， ethics， and durable learning gains.
Ti-5Al-5Mo-5V-1Cr-1Fe （Ti-55511） is a near-β high-strength titanium alloy，which is widely used in aerospace structural components due to its excellent mechanical properties.Benefiting from its high specific strength，good hardenability，and superior damage tolerance，Ti-55511 alloy has been extensively applied in critical load-bearing components such as aircraft landing gear，wing attachments，and fuselage structural parts，where a balanced combination of strength and ductility is required.In this study，the processing-microstructure-property relationship of Ti-55511 alloy was systematically investigated with respect to the effects of different solution temperatures （730 to 790 ℃） and holding times （0.5 to 3.5 h） in the α+β two-phase region on its microstructural evolution and mechanical properties.The morphology and distribution of primary αp phase and secondary αs phase were observed by scanning electron microscopy （SEM）.The results showed that the increase in solution temperature significantly reduced the volume fraction of intragranular primary αp phase，which gradually transformed from a coarse blocky or short rod-like morphology into an elongated lamellar structure.Meanwhile，the volume fraction of the β phase increased，while its stability decreased owing to the dilution of β-stabilizing elements in the β matrix，which provided favorable conditions for the dispersed precipitation of secondary αs phase during subsequent air cooling and aging processes.A higher degree of supersaturation in the β matrix promoted the nucleation of fine secondary αs precipitates，resulting in a dense and homogeneous αs lamellar distribution throughout the grains.The samples subjected to high-temperature solution treatment exhibited a significant increase in the number density of secondary αs，which was distributed as fine lamellae.These fine secondary αs can effectively hinder dislocation movement through the grain refinement strengthening effect，thereby significantly improving the microhardness and yield strength of the alloy.The strengthening effect was primarily associated with the reduced secondary αs spacing and the increased resistance to dislocation slip imposed by more αs/β interfaces.However，the reduced proportion of primary αp at high solution temperatures impaired the deformation accommodating ability of the alloy during tensile deformation，leading to a decrease in elongation.Overall，Ti-55511 alloy exhibited a property evolution phenomenon，in which yield strength increased and elongation decreased with solution temperatures.The present results demonstrated that an appropriate selection of solution temperature and holding time was essential in tailoring the microstructure and achieving an optimal balance between strength and ductility.This study provides valuable guidance for the heat treatment design of near-β titanium alloys used in advanced aerospace applications.
To address the premature failure occurring in the bent region of aerospace-grade 2Cr13Mn9Ni4 austenitic stainless steel unidirectional valve flaps during service， electron backscatter diffraction （EBSD） was employed to systematically characterize the microstructure near the bend radius of specimens bent to 20°， 40°， 60°， 70°， and 80°. The microstructural evolution and deformation-induced phase transformation behavior were analyzed under the coupled gradients of bending angle， through-thickness position， and along-surface position. The results show that the area fraction of the γ phase remained above 90% in all specimens， whereas the area fractions of the ε and α′ phases were relatively low and showed a general increasing trend with fluctuations as the bending angle increased， with local values exceeding 5%. After bending， intragranular orientation spread and local lattice rotation were enhanced. The texture index， kernel average misorientation （KAM）， grain orientation spread （GOS）， and geometrically necessary dislocation （GND） density all exhibited pronounced spatial heterogeneity and did not follow a simple monotonic trend along either the through-thickness or along-surface direction. With increasing bending angle， the peak position of the average GND density shifted from the vicinity of the bend apex toward the tangent point， indicating that the concentrated region of local lattice curvature and equivalent plastic strain gradient changed with increasing bending severity. Orientation relationship analysis supports the γ→ε→α′ transformation pathway. These findings indicate that the local microstructural evolution of 2Cr13Mn9Ni4 during bending deformation is jointly governed by coupled gradients in bending angle， through-thickness position， and along-surface position， providing microstructural evidence for subsequent process optimization and premature failure analysis.
In fully cooperative multi-task multi-agent reinforcement learning， value decomposition methods are expected to exploit shared knowledge across tasks while preserving task-specific coordination patterns. However， when heterogeneous tasks are trained with a single shared mixing network， the common parameter space often suffers from representation conflicts and optimization interference， making it difficult to model task commonality and task differences simultaneously. This issue is particularly critical because the joint action-value function must adapt to varying task characteristics while satisfying the monotonicity constraint required by the individual-global-max principle. To address this problem， this paper proposes MT-MOE， a multi-task value decomposition method based on a mixture-of-experts mechanism. The proposed method is developed under the centralized training with decentralized execution paradigm. Each agent estimates individual action values from local observations， while task information is incorporated into both individual utility estimation and joint-value aggregation. In the mixing stage， a task-aware mixture-of-experts structure is introduced to generate adaptive value-mixing behaviors for different tasks. Specifically， a shared hypernetwork captures common value-mixing patterns across tasks， multiple expert networks provide task-dependent corrective terms， and a soft routing mechanism adaptively combines expert outputs according to task conditions. Through this design， MT-MOE improves the expressive capacity of the joint value function， alleviates negative transfer in multi-task learning， and enhances task-adaptive coordination， while preserving the monotonicity required by value decomposition. Experiments are conducted on the StarCraft Multi-Agent Challenge benchmark， including 9 cooperative maps from three task groups： the SZ series， the MMM series， and the Marine series. All methods are trained for 500K environment steps with 5 random seeds， and performance is evaluated using test win rate. MT-MOE is compared with representative multi-task or transfer-based baselines， including MATTAR， ODIS， and UPDeT， as well as classical value decomposition methods such as QMIX and VDN. Experimental results show that MT-MOE achieves the best overall performance， with an average win rate of 74.90% across the 9 maps. Compared with MATTAR， ODIS， and UPDeT， the proposed method improves the overall average win rate by 9.78， 18.21， and 15.93 percentage points， respectively. Relative to QMIX and VDN， the improvements reach 30.35 and 40.08 percentage points. These results demonstrate that the proposed method effectively enhances the adaptability and generalization capability of value decomposition in heterogeneous multi-task cooperative scenarios. Ablation studies further verify the effectiveness of the proposed design， showing that the multi-expert structure is essential to the performance gain and that adaptive gating contributes to more efficient optimization and more balanced task adaptation. Overall， MT-MOE provides an effective solution for task-aware joint value modeling in multi-task cooperative multi-agent reinforcement learning.
