Data Driven Approach to Optimize the Morphology of Block Copolymer Thin Films

Nanoscale morphology of block copolymer (BCP) films makes them great candidates for nanotechnology, solar cell, membranes and microelectronic applications, where the degree of applicability directly correlates to various morphological properties such as type of morphology, domain spacing, ordering and orientation. However, optimizing morphological properties is a complex process due to the wide range of architectural and processing parameters involved. While Artificial Intelligence (AI) / Machine Learning (ML) tools offer a great potential in streamlining this research, the limited availability of high-quality data presents a significant challenge. Addressing this, we developed a high-throughput feedback model based on morphology characterization data obtained from automated Atomic Force Microscopy (AFM) and Grazing Incidence Small Angle X-ray Scattering (GISAXS) analysis of polystyrene-b-poly (ethylene oxide) (PS-b-PEO) thin films. Our approach leveraged i) robotic platforms—Polybot and Opentrons—for the fabrication and processing of BCP films, ii) high-throughput AFM at Berkeley National Lab’s Molecular Foundry and GISAXS at synchrotron beamlines, ALS and NSLS-II for characterization of films, iii) automated data processing and analysis, and iv) training of data using various ML algorithms. This model used Non-Volatile Additive Solvent Annealing (NVASA), where a high-boiling additive Chloronaphthalene(CN) stayed in the film briefly after spin coating, thereby improving the mobility and ordering of polymer chains. A curated dataset was used to train eleven ML models to predict BCP morphological properties based on processing parameters such as solvent ratio and additive amount (swelling ratio). While the models accurately predicted domain spacing from GISAXS data (R² = 0.84), predictions from AFM data showed lower accuracy due to measurement variability. Despite this, a convolutional neural network (CNN) classified over 93% of AFM images correctly. To improve interpretability, the Shapley Additive exPlanations (SHAP) method was used, revealing that swelling ratio was the most influential processing parameter. This framework provided a foundation for optimizing BCP thin film processing through expanded design space and ML integration. Through establishment of autonomous processing, imaging, and characterization, followed by feedback from ML models, we demonstrate an initial framework for a self-driving laboratory aimed at advancing BCP material design.