Artificial and Augmented Intelligence for Automated Scientific Discovery

AI is a widely used term that conjurers up many of the computers from science fiction. Its stands for a whole collection of ideas, algorithms, computational models and knowledge systems. Recent success of particular types of machine learning (e.g. deep neutral nets) have again excited the interest of the scientific community in delivering insight into the complexity of the real world. This type of approach compliments the knowledge engineering systems that have previously been used, however they require massive amounts of data to be trained. Taking the chemical and materials sciences as exemplar areas we can see that the traditional approaches to scientific discovery work with relatively small amounts of often uncertain data which is distilled by human insight to yield predictions and testable theories which may evolve as new data becomes available. In these areas of science more data is becoming available and the impact of 'larger data' parallels the reality that almost all science now depends on computational assistance. Never-the-less the quantity of quality data needed to train the new AI systems is simply not directly available even with recent advances in automation. As a basis for the network we propose to use 'amplification by simulation' as a key element of the cycle of automated experiments, simulation, AI learning, prediction, comparison, design, further experiments, to create the environment in which leading AI developments can be applied to the chemical and materials discovery.