SceneGenAgent: Precise industrial scene generation with coding agent
Published in: 2025 Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 17847–17875, Vienna, Austria
Xiao Xia, Dan Zhang, Zibo Liao, Zhenyu Hou, Tianrui Sun, Jing Li, Ling Fu, Yuxiao Dong
Generating 3D industrial scenes with LLMs is challenging due to strict requirements for precise measurements and spatial planning. To solve this, the authors introduce SceneGenAgent, an LLM agent that generates industrial scenes via C# code. It uses structured layout planning, verification and iterative refinement to satisfy exact quantitative needs, boosting LLM success rates up to 81.0%. Additionally, the authors release SceneInstruct, a dataset for fine-tuning open-source LLMs like Llama3.1-70B to approach GPT-4o performance within the framework. Code and data are publicly available.
View at publisher's page
SiPhyR: An end-to-end learning-based optimization framework for dynamic grid reconfiguration
Published in: IEEE Transactions on Smart Grid ( Volume: 16, Issue: 2, March 2025)
Rabab Haider, Anuradha Annaswamy, Biswadip Dey, Amit Chakraborty
Rising renewable energy, storage and electric vehicles require faster decision-making paradigms for distribution grids. This paper proposes SiPhyR, a physics-informed machine learning framework for end-to-end learning-based optimization of grid reconfiguration. SiPhyR optimizes topology and power flows to lower line losses, improve voltage profiles and boost renewable integration. To handle NP-hard binary decision variables, it uses a physics-informed rounding method within a differentiable framework. This allows SiPhyR to simultaneously optimize grid topology and dispatch while satisfying safety-critical constraints, as demonstrated on three canonical grids.
View at publisher's page
Leveraging prosumer flexibility to mitigate grid congestion in future power distribution grids
Published in: Energies 2024
Tomaselli D., Most D., et al
The rising adoption of behind-the-meter (BTM) photovoltaic (PV) systems, electric vehicles (EVs) and heat pumps is causing distribution grid congestion. This paper evaluates BTM flexibility using a novel framework that combines a rolling horizon optimal power flow with a piecewise linear cost function to address this issue without costly upgrades. When tested on a grid model of Schutterwald, Germany, the results showed that self-consumption BTM storage resolved all feed-in violations but only 35% of load violations. However, proactively controlling storage charging and discharging effectively resolves the remaining load violations, even in nearly saturated grids.
View at publisher's page
Neural differential equations with integrated control for chemical process modeling
Published in: 2025 IEEE 32nd International Conference on High Performance Computing, Data and Analytics Workshop (HiPCW)
Sumeet Rodiya, Ved Vartak, Kunal Bahuguna, Subhajit Sanfui, Ramsatish Kaluri
Neural differential equations are powerful for chemical process simulation, but their application to controlled chemical processes remains relatively unexplored. This paper evaluates neural ordinary differential equations (NODEs) and neural controlled differential equations (NCDEs) for modeling a batched bioreactor system with explicit control integration. The authors develop a modified NODE framework containing control inputs and study NCDEs, which incorporate continuous control paths using spline interpolation. Implemented in Diffrax for efficient GPU acceleration, the frameworks demonstrate that NCDEs achieve superior predictive accuracy, whereas NODEs offer faster training times.
View at publisher's page
Multi-agent framework leveraging knowledge graphs for virtual commissioning models
Published in: AUTOMATION 2026, Innovation trifft Anwendung (pp. 144-152). VDE
Max Diekmann; Jonas Nitzler; Jan Fischer; Hans-Juergen Pfisterer; Dirk Hartmann
Virtual commissioning models (VCMs) validate discrete manufacturing systems before deployment, but building them is labor-intensive due to fragmented data across tools like Siemens TIA Portal and NX MCD. To address this, the authors propose a knowledge-graph-grounded multi-agent framework. Deterministic tools extract and merge PLC and kinematic data into a unified graph database. A hierarchical agent architecture then assists with system understanding, script-based component generation and cross-domain signal mapping. Tested on a lab-scale manufacturing setup, the approach reduces manual interpretation effort and streamlines early VCM engineering.
View at publisher's page
