R. Holtschneider
Epic Advanced Materials,
United States
Keywords: material characterization, nanomaterials, bayesian optimization, digital twin, multi-modal, multi-objective optimization
Summary:
Boron Nitride Nanotubes (BNNT) are wonder material that will help humanity solve some of the hardest problems of the 21st century like climate change, disease and space exploration and inhabitation. Currently, the impact of Boron Nitride Nanotubes is held back by the lack of scalable, repeatable processes capable of creating high quality BNNTs. Epic Advanced Materials is improving the yield, quality and reliability of the world’s leading process for BNNT production and enabling greater control over the structure and properties of our BNNTs for application specific performance. And we are doing this through an autonomous research loop linking automated multi-modal material characterization analysis, synthesis and experiment selection which is informed by our digital twin [fig 1 in attached pdf]. The properties of nanomaterials are heavily influenced by their nanoscale structure and the properties one application needs may be exactly opposite the needs of another application. To drive improved performance per application, we measure the structural characteristics of our BNNTs such as purity, length, diameter, number of walls, crystalline structure using scanning electron microscopy(SEM),transmission electron microscopy (TEM) and other methods and process parameters such as yield and energy used for each experiment. We have trained machine learning models to identify salient features in these material characterization outputs and post-process these features to receive scores for each characteristic of interest [fig 2 in attached pdf]. These scores are the feedback that drives our autonomous loop. To effectively optimize over a multi-objective space we must determine the optimal tradeoff between objectives in as few experiments as possible. To gain fine grained control over nanomaterial structure, we must understand and incorporate the reaction dynamics of our process. To gain greater insight into the reaction dynamics of our process we have created a digital twin in COMSOL which can determine the locations and temperatures of particles, particle evaporation distance, unit cell pressure and particle velocities.[fig 3 of attached pdf]. To utilize these parameters which have a more direct relationship on the structure of our nanotubes, we have trained a machine learning model on data acquired from simulation which predicts input process parameters we can control physically and in our digital twin such as temperature, pressure, gas flow rates given internal particle dynamics. We then use bayesian optimization methods to explore the internal reactor parameter space to optimize for structural characteristics represented as scores output from our automated analysis. The experiments used to train our Bayesian optimization surrogate model are performed autonomously. The acquisition function of our bayesian optimization determines which internal reactor dynamics should be created next and our machine learning model determines the input parameters required to create those dynamics and instructs our automated reactor to use those parameters in the next experiment. We have used this autonomous research approach to optimize our BNNTs purity and yield, as well as several other structural characteristics for our partners’ specific applications.