ALPHA-Design

The design of new innovative products generally relies on numerical simulations to predict the physical behaviour of the system (mechanical, thermal, electromagnetic, etc.) and carry out its optimisation. These methods, which are well established, are however costly in terms of computing resources and require significant expertise, which hinders their dissemination in industry.

This project aims at exploring an alternative approach, based on neural network learning techniques. Contrary to most approaches in artificial intelligence, the aim is not to approach data but to learn the laws of physics (ordinary differential equations or partial differential equations). The challenge is then to build a neural model including, in a single training, the different physics involved, their couplings, as well as the conditions of optimality of the system.

Such an approach would be a powerful lever for democratising multidisciplinary optimal design and the use of digital twins.