Deep Learning for Manufacturing: an Application to the Rheology Process

Pierre Baqué, Neural Concept

Abstract:

To tackle new challenges, engineers need radically new capabilities, including more effective ways to harness our computational resources. Because of their historical origin, simulation tools are not well adapted to design optimization in fast-paced production and design environments. It is extremely hard for design teams to leverage insights provided by advanced simulations and by the specialised teams who develop them. Deep Learning technologies can be used to integrate knowledge from simulation and design optimization tools in the workflow of the design engineers, instead of delegating this task to separate expert teams. In this talk, we explain how recent algorithms based on Geometric Deep Learning, allow shortcutting any simulation chain through a predictive model that outputs post processed simulation results and optimization suggestions, right from the CAD design. These models are being used in engineering companies to simplify processes and to emulate the expertise of simulation engineers in the hands of product or design engineers early in the development process. Thus, the number of iterations between teams are reduced while accelerating the design activities. In this presentation, we focus on a specific manufacturing application : the rheology process. We will show that the Deep Learning models are used to replace the phase of meshing and solving rheological equations. With inferences based on that model, a classical simulation loop of 4 to 6 days can be shortened to 1 day. In addition, these surrogate models can then be inserted in optimization process to either reduce development time and/or extend the number of optimization iterations to increase performance of the product. The gain in term productivity is up to 30% at iso-perimeter.

Automotive Application Case

An automotive external plastic body part production requires thousands of hours in design, engineering and validations. The mastery of the rheological behavior is a key point for industrial excellence. Calculations are widely used to simulate the material injection in the mold at high pressure and temperature: they determine the part’s behavior in term of mechanical constraint and warpage. The results of simulations are used to set-up molds’ parameters in production and ensure that all different parts can be assembled together. For each vehicle’s part developed by the company from years, many simulation loops have been performed. A large quantity of results, data, know-how have been produced. The presented innovation intends to implement a model estimator based on that knowledge. It is using deep learning methods through 3D geodesic convolutional neural networks (GCNN) to speed-up simulations loops by capitalization over large databases. Historical data and new simulations are gathered in a usable database: 3D geometries, solvers results, injection parameters. This database is used to train a 3D GCNN that can estimate any new geometry. A classical loop of simulation is described as follow: ● The design leader produces a 3D shape with respect to a set of constraints and knowhow recommendations. ● The rheology engineer oversees meshing of the 3D surfaces, prepare the solver parameters with materials choice, injection sequences selection, temperatures, pressure, etc.…, and send it to the solver for computation, ● The solver runs the simulation on CPUs for few hours ● The rheology engineer post processes and takes a decision for an optimization action based on the performance targeted ● The design leader integrates the recommendations and produce a new 3D shape ● The loop continues until an acceptable optimal have been found, in a limited amount of hours allocated Our 3D GCNN models are used to replace the phase of meshing and solving rheological equations. With inferences based on that model, a classical simulation loop of 4 to 6 days can be shortened to 1 day. In addition, these surrogate models can then be inserted in optimization process to either reduce development time and/or extend the number of optimization iterations to increase performance of the product. The gain in term productivity is up to 30% at iso-perimeter. It is important to precise that under no circumstances this model would replace solvers used to calculate the outputs fields. On the contrary, solver will be valorized as their outcome will be encoded in model and used to capitalize any new simulation of a product. The interface for the final user is composed by a simple web browser application: the user provides a new 3D geometry plus a preset of injection parameter, and in less than seconds, the estimated 3D field result is obtained. The project has demonstrated that with only 50 samples, we were able to achieve 82% of accuracy in our results, which was far beyond the expectations. The final users, engineering and development teams are working to enlarge the database and adapt the post-processing steps to add any new simulation result to the database. Once the accuracy of the model will exceed 95% with about 200 simulations in the database, the tool will be ready for use on customers’ projects applications. This innovative method is a precursor. As it has shown good results, it is planned to be deployed and applied to other domains such as crash calculations, static calculations, modal calculations, or thermal simulations. We are also investigating hybrid methods, coupling 3D simulation results with production measurement results. One can say, the era of 3D ANN based data science has begun!