Funded Projects

Data science has been finding its way into the natural sciences and is rapidly emerging as an invaluable research tool. In the chemistry and materials community, data-derived prediction models are now used as efficient surrogates for physics-based models that are traditionally at the heart of modeling and simulations work. The utility of these and other machine learning (ML) approaches is now abundantly clear. However, many ML approaches appear inscrutable. This opacity has rendered them black-box models and hampered the community’s trust in them. The new field of ‘eXplainable Artificial Intelligence’ (XAI) sets out to develop techniques that can explain AI/ML models. XAI aims to create visibility into the predictions and decision-making processes of AI/ML system, explain the rationale behind the decisions, and illuminate the strength and weakness of these processes. XAI is still a novel approach in the data science community, and it has barely been explored in application domains such as chemistry. Given the mistrust in AI/ML, we believe that advancing XAI techniques that can interpret and explain ML models for molecules, materials, and reactions (i.e., making the inherent working of chemical ML models comprehensible to research practitioners and allowing them to derive meaningful insights) is vitally important for the sustained success of ML in chemistry and materials. XAI will alleviate the dangers of making data-derived predictions that are devoid of a robust scientific basis. The ability to explain the results of ML models in chemistry will not only add scientific value, but will also help improve our understanding of the integral relationships between the structure of a compound and its properties. The information provided by XAI can further be leveraged to improve a given model via physical constraints, feature engineering, hyperparameter tuning, modifying its architecture, and/or tailoring training data sets in active and transfer learning approaches.