Replicating TRIZ Reasoning Through Deep Learning

For two decades, TRIZ has been considered as an inventive approach without rival in the existing design methods. It owes its originality to the work of Altshuller and his colleagues who compiled a large amount of scientific and technological data from all domains to build generic meta-models that inspire its users. But in its history, TRIZ has also met detractors who point out above all its learning complexity and the lack of scientific rigor of its description. This article presents the progress of our research in the use of Artificial Intelligence and in particular the progress made in reproducing TRIZ reasoning through the Deep Learning approach on a large quantity of trans-disciplinary patent sets. We describe the approach used, propose and discuss two case studies that artificially reproduce TRIZ reasoning in order to test the relevance of such an approach and its perspectives for the future of our research.