Machine learning can emulate scientific modeling and be incorporated into real-time process controls.
Machine learning can emulate scientific modeling and be incorporated into real-time process controls.

Engineers at Vitro Glass, in partnership with high performance computing (HPC) experts at Lawrence Livermore National Laboratory (LLNL), have created a machine-learning algorithm that allows the company to make real-time process control adjustments to furnace operations.

The surrogate model, an HPC-developed machine learning tool, is one example of how supercomputing resources and expertise can improve the energy efficiency, productivity, and competitiveness of U.S. manufacturing.

Plate glass, commonly used in windows and doors, is produced in furnaces that melt raw materials, strip gases and impurities, and homogenize the glass. Small variations in this process can lead to temperature disruptions that affect the efficiency of glass manufacturing and lower the quality of the final product.

Glass engineers have traditionally used complex Computational Fluid Dynamics (CFD) models to observe these disruptions and devise strategies to return to normal production. This process can take up to two weeks, during which the furnace cannot produce a viable product.

In 2017, engineers at Vitro Glass began working with researchers at LLNL to use HPC to reduce the time required to correct a furnace’s operational parameters in the event of a process disruption. Through the High Performance Computing for Manufacturing (HPC4Mfg) program, LLNL proposed using the glass manufacturing CFD model to generate a fast-running surrogate. The machine-learning surrogate is much less computationally intensive than the original CFD model and can be used as a real-time furnace control system, running on a desktop computer in minutes rather than weeks. LLNL supercomputers made it possible to run the simulations needed to generate the surrogate in a matter of weeks rather than years.

To build the surrogate model, Vitro Glass identified a range of parameter settings of interest to its glass furnace. Then, researchers ran hundreds of CFD simulations covering that range of settings to produce detailed predictions of the operating conditions in the furnace. The data from these simulations, combined with machine-learning techniques, allowed researchers to train a neural network capable of reproducing the simulation data in the range of furnace parameters that Vitro Glass identified. As a result, the algorithm effectively “learns” how the system responds to an operator’s adjustment to process parameters or sensor-reading changes.

The new control system acts as a fast-running prediction tool that based on machine learning can instantly tell the furnace operator how to adjust when conditions deviate from normal. This fast-running prediction tool can save roughly two weeks of production per furnace every year. The prediction tool can also reduce the amount of substandard glass that must be discarded, increasing glass-manufacturing productivity by 2%. If implemented broadly, these improvements could save the U.S. glass manufacturing industry approximately 2.5 TBTUs of energy per year, nearly 4% of the energy used to produce flat glass.

Through the HPC4Mfg program, this technique for process control is being applied to other industrial-control scenarios, including aluminum casting and steel sheet rolling.

The Office of Energy Efficiency and Renewable Energy’s Advanced Manufacturing Office funds the HPC4Mfg program, which is a part of the High Performance Computing for Energy Innovation (HPC4EI) initiative. In addition to manufacturing, the initiative includes subprograms for materials and mobility.

Individual projects receive $300,000 and access to the HPC resources and expertise of DOE’s National Laboratories. Industry partners in these initiatives share at least 20% of the cost. The program accepts project proposals twice per year.

For additional information, visit the HPC4EI website.