Virtual product development using design simulation is an effective strategy to meet the evolving needs of the product development process. Design simulation involves modeling and simulating real-life scenarios to virtually test product performance. Various software packages that utilize either finite element (FEA) or computational fluid dynamics (CFD) methods are used for design simulation. Depending upon the complexity of the physical problem being solved, these simulations require time and computational resources. Our ability to perform high-end analysis- FEA or CFD, without worrying about upfront hardware infrastructure costs, our ability to share data and collaborate with globally dispersed teams has become a reality using Dassault Systèmes’ 3DEXPERIENCE platform.
While product design simulation for virtual product testing helps us identify design flaws early in the design phase, any modification in design would require the analysis (i.e. design simulation) to be solved again. Parametric simulation setups and automation of simulation processes help expedite the virtual design validation. However, these techniques require the user knowledge, time, and computer power. This is where Artificial Intelligence using Machine Learning can be leveraged.
Machine learning is a way to train computer systems by feeding them large sets of curated data. It learns the data by discovering complex patterns and relationships between inputs and outputs. The computer gets trained to perform specific tasks and generate useful insights from datasets. These newly gathered insights can then be applied to related new scenarios to predict the outcomes. By letting computers handle the heavy lifting of performing repeatable tasks, we can have more time for creative problem-solving.
Although iterations of this technology have existed since the 1950s, recent advances in data collection and computing power have exponentially increased the abilities of this disruptive technology. In the realm of product design, the integration of machine learning (ML), particularly neural networks, into physics simulations is revolutionizing how designers and engineers approach design simulation.
Neural networks, inspired by the human brain’s structure, are a subset of machine learning that excel at recognizing patterns and making predictions from large datasets. It consists of many simple processing nodes – densely interconnected neurons. An individual node might be connected to several nodes in the layer beneath it, from which it receives data, and several nodes in the layer above it, to which it sends data [1]. These individual nodes continually recalibrate themselves with each sample of training data so that input data correlates with the correct output.
When applied to structural design simulations, neural networks offer significant advantages such as accelerated computational speed and freeing up resources for other important tasks. To illustrate the process, a linear static analysis was performed using SOLIDWORKS Simulation to evaluate stresses in press assembly subjected to force. This served as the baseline model. Several simulation studies were conducted varying the dimensions of the components to generate training data. A parametric study was also used to speed up the process of generating training data.
Eight percent of the generated simulation data was used to train a feedforward type of neural network using an open-source code. Twenty percent of the simulation data was used for testing the accuracy of the neural network to predict output. Only numeric values were used for training purposes. A comparison of stress values for the trained model and test model is as shown below.
For linear static analysis, the model predicted the stress values with a mean absolute percentage error of 0.2535 and an R-square (coefficient of determination) value of 0.993.
The workflow was also tested for a nonlinear analysis using the Structural Mechanics role on the 3DEXPERIENCE platform. The top loading of a plastic bottle was simulated using nonlinear dynamics analysis that applied a prescribed displacement and evaluated reaction force. The training data for the neural network was generated by solving several simulations that calculated the reaction force for various thickness values and shapes of the bottle.
Only numeric values were used to train a feedforward neural network with a backpropagation algorithm. As shown below, the neural network consisted of that consisted of two, three-input nodes – thickness, displacement, and shape factor, one output node for reaction force and two intermediate (hidden) layers.
Twenty percent of the simulated data that was not used in the neural network training process was presented to test the prediction accuracy of the newly trained model. For this case, the R-square value was 0.725 with a mean absolute percentage error of 0.386.
While integration of neural networks in structural simulations offers numerous benefits, it also presents challenges. These include the need for large, high-quality datasets and the complexity of training accurate models. However, ongoing advancements in neural network algorithms and computational power are steadily addressing these issues.
This approach has enormous potential to boost productivity, reduce the time to market, and drive innovation in product design. By harnessing machine learning in structural simulation, product designers and engineers can efficiently create more robust and optimized products, meeting the ever-growing demands of modern consumers and industries.
