Conference Agenda

Discrete manufacturing industry faces multiple challenges due to data silos arising from the use of multiple Management Information Systems (MIS) in the company leading to poor cross-system traceability and make root-cause analysis of a failure or parts/components manually intensive and prone to data inconsistencies. Traditional relational databases are not capable of mul-ti-hop queries required for complex industrial troubleshooting, such as link-ing a field failure back to specific design tolerances, machine or robot effi-ciency and supplier batches. Large Language Models (LLMs) are able to of-fer intuitive interfaces with reasoning capability but standard Retrieval-Augmented Generation (RAG) often encounters hallucinations due to a lack of context and awareness regarding Product Development, Manufacturing and Supply Chain Management (SCM) schemas. This research proposes a novel GraphRAG (Graph Retrieval Augmented Generation) based orchestra-tion framework designed to bridge these gaps by mapping siloed data into a Unified Knowledge Graph (UKG) of Product Lifecycle Management (PLM), Enterprise Resource Planning (ERP) and Manufacturing Execution Systems (MES). The Systems Architecture (SA) utilizes an LLM orchestrator which acts as a central hub for managing the flow of Data between the User, LLM and UKG.

This study presents a pipeline that combines Large Language Models (LLMs) and Knowledge Graphs (KGs) to automate the extraction and validation of semantic features from technical requirements in industrial engineering documents. The proposed approach focuses on identifying physical quantity names, numerical values, and units of measurement from natural-language requirements and transforming them into structured, machine-readable outputs suitable for integration with Product Lifecycle Management (PLM) systems. A LLaMA 3.1 8B model was fine-tuned on a manually annotated synthetic dataset and compared with GPT-4o as a general-purpose baseline model. The extracted information was further validated through the Quantities, Units, Dimensions and Types (QUDT) knowledge graph, which supports consistency checking between units of measurement and physical quantities. The fine-tuned model outperformed GPT-4o in complete requirement-level extraction, achieving an Exact Match (EM) score of 73.7 after knowledge graph integration. A case study involving a pressure-regulating valve further demonstrated the applicability of the complete pipeline to a different technical domain, achieving an EM score of 90 and enabling the creation of structured requirement objects in a PLM environment. The results indicate that combining task-specific LLM fine-tuning with knowledge graph-based validation can support the automated formalization of technical requirements and reduce the manual effort required to transform unstructured documents into machine-readable engineering information.

Recent advances in large language models (LLMs) have created new opportunities for supporting engineering innovation and inventive problem solving. However, current AI-based ideation systems still exhibit important limitations, including weak integration between creative generation, prior-art analysis, novelty assessment, and iterative refinement within unified engineering-design workflows. This study presents Avolution, a retrieval-grounded multi-agent framework designed to support systematic innovation through the integration of TRIZ-inspired reasoning, engineering-design methodologies, retrieval-augmented generation, and generative AI. The proposed architecture coordinates specialized agents dedicated to problem reformulation, prior-art retrieval, inventive-trigger generation, exploratory concept generation, and design-around refinement. The workflow was implemented as a modular orchestration pipeline on the n8n platform, integrating patent repositories, scientific databases, web retrieval systems, and multimodal generative modules within a unified human-in-the-loop environment. Experimental validation was conducted across heterogeneous engineering-design scenarios involving compliant mechanisms, energy harvesting systems, and safety-critical transportation problems. Compared with unconstrained monolithic LLM ideation, the proposed framework expanded the explored solution space, improved technical grounding and prior-art awareness, and generated more contextually relevant and structurally differentiated engineering concepts. The results suggest that orchestrated retrieval-grounded multi-agent reasoning may represent a promising direction for supporting systematic engineering innovation beyond conventional generative-AI interaction paradigms.

Airport surface congestion is expected to intensify over the coming decades due to sustained growth in air traffic demand and the limited expansion of airport infrastructure. This paper proposes a hybrid taxi queueing model to estimate the portion of taxi-out delay attributable to traffic-induced conges-tion at saturated airports. The model is subsequently used to evaluate the im-pact of non-stop taxiing implementation on taxi-out times under congested operating conditions. The objective is to assign a normalized value, ex-pressed in terms of perceived effectiveness for airline operators, associated with the implementation of local 4D taxi trajectory optimization solutions. Results obtained from two airport pilot cases indicate that local optimization strategies can provide significant benefits at congested but not fully saturated airports. In highly critical conditions, however, such solutions should be complemented by measures aimed at reducing overall system variability. These findings highlight the importance of integrating taxi routing optimiza-tion with complementary surface traffic management strategies to fully un-lock system-level operational benefits.

A 3D finite element modelling workflow for bicycle tyre analysis is presented, addressing the limited use of detailed numerical simulations for bicycle tyres compared with the automotive tyre sector. Unlike most tyre FE models relying on axisymmetric geometries or simplified tread representations, the proposed workflow explicitly includes the tread pattern of an off-road bicycle tyre within a full 3D model, together with bead reinforcement and body ply structure. The workflow is implemented in the implicit LS-DYNA solver and reproduces tyre mounting, inflation, and quasi-static vertical loading. The approach is applied to a 29 × 2.40 off-road tubeless bicycle tyre and evaluated in terms of footprint characteristics and vertical stiffness under different inflation pressures. The simulations qualitatively reproduce the characteristic elliptical footprint shape and capture the stiffening effect associated with increased inflation pressure. A vertical stiffness of 69.6 N/mm is predicted at an inflation pressure of 0.15 MPa, showing reasonable agreement with experimental data available in the literature. Nevertheless, the predicted footprint dimensions are smaller than reported experimental measurements, highlighting the need for dedicated material characterisation and tyre-specific validation before the model can be used for fully predictive design applications.

High-fidelity simulations of lithium-ion batteries are often limited by the high computational cost associated with fully resolved three-dimensional (3D) multi-physics models. This work presents a hybrid 3D–1D framework for simulating a large-format 4680 cylindrical lithium-ion cell by coupling a one-dimensional electrochemical model with a three-dimensional thermal representation of the jellyroll. To ensure the physical reliability of the core governing equations, the baseline 3D fully solved solver architecture was successfully validated against recent experimental calorimetric data from a standard 18650 cylindrical cell, demonstrating a maximum relative error below 3% and proving grid independ-ence through a systematic mesh sensitivity study. The validated framework was then implemented in COMSOL Multiphysics to compare the fully resolved 3D model with the proposed hybrid approach under constant-current constant-voltage (CCCV) charging and constant-current (CC) discharging at C/2, 1C, and 2C. The hybrid model accurately reproduces the cell's thermal evolution, exhibit-ing relative errors below 0.5% at C/2, approximately 2% at 1C, and around 5% at 2C. Concurrently, the computational demand is drastically reduced from approx-imately 2.3×10¹⁹ to 4.1×10¹⁶ floating-point operations, thereby confirming the effectiveness, accuracy, and scalability of the proposed hybrid framework for efficient multi-physics battery simulations.