Throughout history, humans have harnessed matter to perform tasks beyond their biological limits. Initially, tools relied solely on shape and structure for functionality. We progressed to responsive matter that reacts to external stimuli and are now challenged by adaptive matter, which could alter its response based on environmental conditions. A major scientific goal is creating matter that can learn, where behavior depends on both the present and its history. This matter would have long-term memory, enabling autonomous interaction with its environment and self-regulation of actions. We may call such matter ‘intelligent.’
In this lecture, we introduce a number of experiments towards ‘intelligent’ disordered nanomaterial systems, where we make use of “material learning” to realize functionality. We have earlier shown that a ‘designless’ network of gold nanoparticles can be configured into Boolean logic gates using artificial evolution. We later demonstrated that this principle is generic and can be transferred to other material systems. By exploiting the nonlinearity of a nanoscale network of dopants in silicon, referred to as a dopant network processing unit (DNPU), we can significantly facilitate handwritten digit classification. An alternative material-learning approach is followed by first mapping our DNPU on a deep-neural-network model, which allows for applying standard machine-learning techniques in finding functionality. We also can optimize DNPUs by using gradient descent in materia, using experimental gradient extraction. Finally, we show that our devices are not only suitable for solving static problems but can also be applied in highly efficient real-time processing of temporal signals at room temperature.