(370l) Semi-Automatic Method to Optimize Multi-Lamp High Flux Solar Simulators Utilizing Machine Learning Algorithms | AIChE

(370l) Semi-Automatic Method to Optimize Multi-Lamp High Flux Solar Simulators Utilizing Machine Learning Algorithms


Kakosimos, K. E. - Presenter, Texas A&M University at Qatar
Al-Hashimi, M., Texas A&M University at Qatar
Arshard, A., Texas A&M University at Qatar
Hafeez, S., Texas A&M at Qatar
Hassan, M., Texas A&M University at Qatar
Khalil, B., Texas A&M University at Qatar
Extensive research is being conducted on photon energy driven reactions which include wastewater treatment, H2 production, and advanced material aging under diffuse or concentrated light. Solar simulators are a convenient tool for the latter research applications because they allow emulation of solar radiation under well controlled laboratory conditions.

Most concentrated light simulators are comprised by multiple light sources with ellipsoidal or parabolic mirrors. Thus, they require accurate characterization of various system parameters such as peak flux and flux density distribution. At the same time, they also require optimization to ensure system can operate at its theoretical maximum flux and provide the necessary radiating energy. However, this process is either manual or semi-automate and demands expert-user intervention – i.e. a tedious and time consuming process.

This study utilizes the High Flux Solar Simulator (HFSS) facility at Texas A&M University comprised by seven short-arc Xe lamps of 6 kW each. We present an automated system for the characterization of the irradiance, collection of experimental data, and control of the irradiance. At first, irradiance was characterized by using the flux mapping method. In this method, the greyscale value from illuminated target image is correlated to flux gage data to obtain a calibration curve. The target images were normalized for several exposure times. The data acquisition was automatic and included image capture, target movement, etc. The collected data was processed using an in-house algorithm for the calculation of the flux parameters. Then, the data were used to train a semi-supervised machine learning algorithm based on a typical convolutional neural network model. Finally, the model was used to optimize alignment of the light sources (three degrees of freedom per source) given variable flux parameters i.e. peak flux and flux distribution. The proposed methodology is expected to facilitate initial deployment of high flux solar simulator. It will also assist on the dynamic control of reactor conditions i.e. emulating variable overcast or daily sunlight variability.