AMP_06_September_2021

A D V A N C E D M A T E R I A L S & P R O C E S S E S | S E P T E M B E R 2 0 2 1 2 1 M aterials informatics has matured over the last five years, spread- ing from university research labs to leading multinational materials and chemicals companies. The knowledge surrounding what it can and can’t do is varied—some scientists are skeptical of the “black box” of artificial intelligence (AI), while others buy into the hype that AI performs alchemy in R&D. This article outlines what AI can do and why businesses are using it, illus- trated by examples. First, it is important to acknowledge the challenges of ap- plying AI to materials development. CHALLENGES IN MATERIALS INFORMATICS Materials data is small data; each data point requires a laborious experi- ment. It is also complex, requiringmeta- data on test conditions and a full pro- cess history. Lastly, the data comes in many different formats: chemical for- mulas, microstructure images, curves, and so on. It is perhaps for this reason that materials science has been slow to take advantage of the benefits of AI. Over the last decade, researchers have been re- fining data models, data ingesters, and AI models to tailor them to be used on materials (Fig. 1). That work is now bearing fruit. Rather than putting people out of jobs, AI for materials relies on the do- main expertise of materials scientists. If only a small dataset is available, it is critically important to use all the in- formation the team has to improve the results of the AI model. Scientists can incorporate their domain knowledge in an AI workflow through: • materials-specific featurization • diverse data selection • model configuration and algorithm selection • design space definition • linking simulation and/or analytical equations with AI Scientists are often hesitant to trust a black box, but they are con- stantly looking to learn new things. It is therefore critical for them to under- stand why an AI model is making the predictions it is and how uncertain its predictions might be. More interpretable AI platforms make it possible to see which input data is having the most effect on the predic- tions (feature importance). In this way, scientists can check that the model is performing properly and occasionally learn something new about how the un- derlying science is working. Displays of measures of model ac- curacy and plots of predicted versus ac- tual values using validation data also help scientists understand what’s go- ing on behind the scenes. Crucially, pre- dicted values are listed as a value with a quantified uncertainty so that scientists understand how much weight to attach to the data. AI does not have to be a black box, and it is not taking over scientists’ jobs. AI FOR MATERIALS DEVELOPMENT IS COMING OF AGE Despite challenges, research and development scientists are increasingly turning to artificial intelligence when creating new materials. TECHNICAL SPOTLIGHT Embracingmaterials informatics can enable thematerials community to speed up growth.