An exhaustive guide to Explainable AI for Material Science & Informatics

High-performance Machine Learning / AI algorithms have been inadequate in precisely explaining their so-called accurate predictions. In the majority of cases, there are fundamental issues ( overfitting, high complexity, generalization errors ) with the models and/or the process that haunt engineers later, after they’ve invested a significant amount of time and resources, eventually finding out the discrepancy. Our team of experts ( Polymer and AI Scientists ) solved this problem by using a comprehensive closed-loop strategy. The AI engineers made sure they open the black-box models (with explainability and interpretability), to quantify and highlight relationships between the input formulations and output properties in the data. Consequently, Polymer Scientists were responsible for analyzing and validating given relationships with academic research and theoretical frameworks of the industry before approving the model for usage. This synchronous balance of engineering and domain expertise helped us build consistent, robust, and domain-expertise-backed models that drive real growth for R&D Labs across the world. It also instills trust and confidence in our customers as we provide explanations for our optimized recommendations to achieve desired objectives/properties. In this white paper, we will be sharing a practical, solution-centric application of Explainable AI for Material Synthesis using a state-of-the-art, game-theory-based algorithm: SHAP. In addition to code and fundamental guidelines, we will also be sharing our insights and experience in building and enhancing the engine over time.

INTRODUCTION

Artificial Intelligence has empowered and transformed domains and industries with powerful data-based decisions. With an expanding influence and diverse applications in Healthcare, Pharmaceuticals, Marketing, Finance, etc, AI is responsible for opening up a world of opportunities for businesses to lead the future. The primary facilitators for this non-linear growth are access to a lot of data and cheap computation power.

While the combination of computational power and data has fueled the growth of AI, there still exist numerous use cases where AI hasn’t been used extensively. One such arena is Material Science and Informatics. Material Sciences refer to the study of materials that gives us the ability to produce materials that meet the ever-growing market demands. Whether you wish to build encapsulations for batteries of an electric car, lightweight polymers for Mars rovers, or water-resistant paints for submarines, Material Science is to the rescue!A pressing problem with Material Sciences is the critical reliance on expensive, time-consuming experiments to achieve desired material behavior. These experiments involve a lot of variable parameters ( chemicals / solvents / catalysts / processing conditions ) and corresponding properties ( elongation / compression / heat sensitivity ) with complex interactions and interconnected effects. The goal is to understand the underlying relationships and conduct optimized experiments to reach the desired objectives faster and cheaper.

AI has the ability to learn patterns from historical data and generate accurate predictions on new, unseen data. By leveraging the data from existing experiments, we can build models to predict the outcomes of new experiments which potentially reduces the burden on researchers as their search space is narrowed down with a decrease in cost and resource efficiency to produce these materials. Correspondingly, we can also use the learned relationships to generate recommendations, the combination of inputs that achieve the desired property. But there’s a catch!

“All models are wrong, some are useful”

• George E. P. Box, Statistician

When we envisioned the AI Engine in the Polymerize Lab platform, we explored and exploited strategies for AI models to be in continuous sync with domain experts and constantly exchange valuable insights with scientists and researchers. That is where we introduced more than just AI, Explainable AI. After extensive research on best performing AI / Machine Learning algorithms, we also explore the explainability of these models. Eventually, we found that a secret jackpot lies hidden, under the hood - in the complex mathematical equations. Therefore, it was obvious for us to provide our customers with additional value and deeper insights into their experiments. So we opened the black box and rejected the lossy trade-off of explainability to higher performance and opted for the best of both worlds. We made sure to find the quantitative and qualitative relationships ( using multiple plots and figures ) and validate them ( with domain expertise and research ).

‍