COGNITWIN

We say that an Artificial intelligence (AI) is able to construct "world models" when it is able to develop a model of the surrounding physical environment and, most importantly, to foresee the consequences of its actions on it. For an AI to develop models of complex physical systems, we hypothesize that it is necessary to equip it with inductive biases that should be precisely those based upon the fulfillment of known physical principles. Incorporating existing physical knowledge in the form of inductive biases seems to be the best way to: 

• avoid unexpected results in previously unseen situationsi.e., extrapolation, 
• maximize accuracy in the predictions 
• minimize the amount of data necessary at the training stage. 

The limitations faced by industry in the use of the techniques that today constitute the state of the art are several: the very high computational cost of current techniquesalso linked to energy consumption and sustainability, the very high specialization of the personnel who must handle them or the lack of experimental information on the materials, processes or physical laws that are applicable, among them. 

To overcome these limitations, COGNITWIN proposes to develop a new generation of physics-informed hybrid artificial intelligence techniques that will enable a substantial advance in terms of a) AI sustainability, b) AI Efficiency, c) Reliability of AI, d) Ethics of AI, e) Human-centred AI. 

COGNITWIN will therefore develop the disruptive concept of hybrid AI, which will unite data-driven models with those based on wellestablished physical laws. The hybrid paradigm allows for a reduction in the volume of data used, while achieving a substantial approach to reliable AI by ensuring that physical laws are adhered to in predictions. Depending on the application, hybrid AI can be the decisionmaker, with humans acting as supervisors of the AI, or, conversely, humans can be the decision-maker, with hybrid AI assisting in the decision-making process.