Published in Engineering, this research from PolyU and Guangzhou Institute of Industrial Technology develops a ConvLSTM-Fire deep learning model to repair broken sensor data and reconstruct 2D building temperature fields. Paired with a ChatGPT-4o-based LLM self-driven agent, the hybrid system converts real-time fire thermal data into actionable emergency guidance. Quantitative tests prove full 2D temperature inputs greatly improve the accuracy of LLM firefighting decisions, supporting voice-aided on-site response for occupants and firefighters.

Published in Engineering, this study introduces HCGCDM, an LLM-powered cognitive diagnosis model built on heterogeneous concept graphs. Unlike traditional tools that only capture prerequisite knowledge links, it automatically mines three types of concept relations via RAG. Tailored graph encoders and adaptive weighting boost student proficiency assessment accuracy, with ablation and LLM comparison experiments verifying its effectiveness and the impact of LLM capability on diagnostic performance.

Published in Engineering, this report introduces LLM-IDA, a three-tier multi-agent LLM framework for automated industrial design. It divides design workflows into four functional modules to address LLM hallucination and poor robustness via closed-loop code generation and self-verification. Benchmarked against RAG with pass@10 metrics and ablation tests, the modular system cuts manual engineering overhead and supports cross-domain mechanical design optimization, with its current limits and future research directions covered.

Published in Engineering, this news report covers a comprehensive survey exploring how large language models tackle long-standing flaws in two mainstream autonomous driving frameworks: modularized pipelines and end-to-end neural networks. It details LLM applications across perception, prediction, planning and data construction, lists core deployment barriers like hallucination and latency, and unpacks ongoing debates over AGI and the future architecture of intelligent driving systems.

Traditional manual design of Earth Observation Satellite Scheduling Problem algorithms costs massive expert labor. This research introduces AgentAD, an LLM-powered five-agent collaborative framework that converts natural language scenario descriptions into executable Python scheduling codes via four sequential phases. Tests prove its self-evolved algorithms outperform mainstream human-designed heuristics, with memory stream and chain-of-coding mechanisms stabilizing iterative optimization for both basic and complex satellite task scenarios.

Tired of slow, inflexible wind farm maintenance scheduling? New research published in Engineering presents LLM4M, a dedicated fine-tuned language model for turbine repair planning. It quickly generates cost-efficient crew routes with tiny errors versus mathematical optimal solutions, beating popular general LLMs. It runs offline on ordinary hardware and adapts to dynamic real-time maintenance tasks, while its discovered performance phase transition offers practical tips for training industry-specific AI models.

Wildly conflicting global energy estimates for data centers have long blocked reliable low-carbon planning, with 2020 power use figures varying sixfold. This new research in Engineering breaks down flaws in two mainstream calculation methods, introduces AI load recognition for unregistered server hubs, and calls for standardized grid-linked energy registration. It also reveals untapped temporal and spatial load flexibility of AI computing facilities to balance renewable power supplies, offering actionable technical and policy fixes for clearer data center energy governance.

Published in Engineering, this review systematically sorts interdisciplinary engineering therapies for spinal cord injury. It analyzes the three-stage secondary pathological cascade after trauma and elaborates five core intervention branches including neuromodulation, cell transplantation, and functional biomaterials. The article also identifies major translational obstacles and advocates integrated multimodal treatment plus cross-field cooperation to speed clinical transformation of SCI repair technologies.