Vivix builds AI digital twin of its glass furnace with Siemens, Snowflake, NVIDIA

A furnace that is hard to observe

Vivix makes glass and mirrors for the construction and furniture markets. In float-glass production, the furnace is one of the most critical assets and one of the hardest to observe, according to the post. At operating temperatures of approximately 1,550°C to 1,600°C, teams have limited direct visibility into how combustion gases, molten glass and the refractory walls interact.

Understanding those conditions while production continues is essential for decisions about energy use, asset life and glass quality. The xDT combines engineering simulations, observed furnace conditions and AI models to help teams estimate conditions they cannot measure directly and to evaluate possible operating changes. Vivix is focusing first on three connected areas: refractory health, combustion efficiency and glass-melting quality.

How the architecture works

The post describes a trade-off: traditional physics simulations give high-fidelity predictions but are computationally intensive, while standard AI models often lack an understanding of physics. The xDT tries to avoid that compromise. High-fidelity Siemens Simcenter multiphysics models are converted through Simcenter reduced order modeling into lightweight, portable executables and deployed natively within Snowflake's AI Data Cloud. The Simcenter Executable Digital Twin Manager handles the xDT lifecycle in a single solution.

Snowflake's role is to put the engineering insights next to Vivix's operational data. Vivix brings IT and operational technology data, including physical sensor readings, together in a unified namespace in Snowflake. Teams get a governed view of observed furnace conditions that they can use alongside the twin's predictions, instead of reviewing simulation results in isolation.

NVIDIA PhysicsNeMo lets Vivix build AI surrogate models that combine physics-driven causality with simulation and observed data for real-time predictions. Technologies built using NVIDIA Omniverse libraries let engineering and operations teams review high-fidelity results together. The solution runs on Amazon Web Services (AWS).

The result works as a continuous "virtual sensor". It reconstructs internal states that cannot be measured, including volumetric heat distribution, gas velocity fields and thermal fluxes across refractory walls, without a constant connection to heavy simulation environments.

Three areas of optimization

• Refractory preservation: by fusing real-time thermocouple streams with refractory thermal profiling, the twin calculates localized wear rates and detects emerging hotspots hours in advance. The aim is to extend furnace campaign life before cold repairs are needed.
• Combustion and energy efficiency: airflow and thermal models evaluate burner power levels and heat distribution patterns. This supports setpoint optimization to stabilize combustion, lower natural gas consumption and reduce emissions.
• Glass melting and quality: modeling thermal gradients and fluid flow in the molten glass bath gives visibility into residence time distribution and melting completeness. This is meant to improve glass homogeneity and reduce defect rates.

Based on the pilot project, Vivix expects to cut furnace energy consumption by 3%-5%, extend furnace lifespan by one year and shorten its data collection and analytics cycle by 80%. These are expectations stated in the post, not results from full production.

Natural-language access and what it means

Combining Siemens Simcenter physics models with Snowflake Cortex AI, the hybrid model can predict furnace behavior under new conditions and run virtual "what-if" simulations. A Snowflake AI multiagent interface lets operators talk to the twin in natural language. Engineers can ask about optimal air velocities, test scenarios and receive root-cause diagnostics without risking the physical asset.

For heavy-industry operators, the case shows one way to bring simulation, sensor data and AI into a single governed environment. Vivix says that testing changes on the virtual twin, rather than putting physical assets in unknown and potentially unsafe operating conditions, gives it a safer and more predictive foundation. The figures so far come from a pilot and the company's own projections.