Message-passing neural networks for high-throughput polymer screening

Machine learning methods have shown promise in predicting molecular properties, and given sufficient training data ML may surpass density functional theory in computational speed and chemical accuracy. However, the most accurate machine learning methods require optimized 3D molecular geometries, limiting their applicability for high-throughput screening. We show that near-optimal results for large polymeric molecules can be obtained without optimized 3D geometry, and that trained model weights can be used to improve performance on related tasks.