RUDN University professor and colleagues from Italy, Canada, and Turkey built a deep neural network that predicts the strength of composite materials after processing with almost 100% accuracy. The results were published in the first quarter of Materials.
Newswise — Metal matrix composites are a modern alternative to steel. It is a “molecular-reinforced material”, which consists of a metal reinforcement matrix and filler. Such composites can be further strengthened, for example, using titanium monobromide or titanium carbide. But even such materials are not immune from destruction, so they are additionally processed by shot blasting: a powerful gas jet with small particles makes it stronger. To select the optimal parameters for shot blasting, an extensive experimental base is needed. However conducting such experiments is expensive and time-consuming, and existing simulators do not give plausible results. The RUDN University professor and colleagues from Italy, Canada, and Turkey trained a deep neural network for this.
“Most failure in metallic materials starts at the surface. This also happens in critical parts of industrial production. Therefore, shot peening can play a decisive role and improve the mechanical properties of the surface - hardness, resistance to corrosion and wear. Testing in this area is labor-intensive and expensive. Simulation programs contain many errors, which can lead to catastrophic damage,” Kazem Reza Kashyzadeh, professor of the RUDN Department of Transport said.
To train the neural network, the researchers used experimental data from shot-peening titanium composites with different contents of reinforcers. The RUDN University professor took the content of amplifiers - monobromide and titanium carbide - and the intensity of processing as input data. At the output, the neural network describes the hardness and residual stresses throughout the entire depth of the material.
The accuracy of the constructed neural network turned out to be almost perfect. The material hardness is predicted with 99.4% accuracy, and the residual stress with 98.8% accuracy. Previous results obtained with a shallow neural network were about 1% worse. A neural network helps to understand how the intensity of processing will affect the result. For example, the most pronounced residual stress appears at a depth of up to 15 micrometers from the surface at an intensity of 0.25-0.30 on the Almen scale.
“The predictions of the deep neural network were 0.98% more accurate than those of the conventional one. Therefore, deep neural networks can be considered a powerful tool for analyzing hardness and residual stress after shot peening,” Kazem Reza Kashyzadeh, professor of the RUDN Department of Transport said.
RSS