Abstract
To date the environmental impacts of production of most chemicals are unknown, hindering sustainable process evaluation and design. Life cycle assessment (LCA) is the most comprehensive method for estimating these environmental impacts but existing LCA databases cover a small fraction of the industrial chemicals. Environmental impact prediction models promise to bridge this gap; several models were published, with the apparent limitations of limited accuracy and applicability. The very small number of available LCA datasets for industrial chemicals is the major issue for developing environmental impact prediction models; this limits the domain of applicability and accuracy of the models. Here we show that a machine learning approach using transfer learning partially addresses these limitations. We found supervised pretraining on molecular price data increases model accuracy for three different deep learning model architectures, for two out of three different environmental impact categories. Taking the mean for model architectures, across all impacts, transfer learning increases
R
2
by 0.11, and reduces mean percentage absolute error by 2.5%. Our most accurate model for carbon footprint prediction achieves an
R
2
of 0.62, similar to existing literature benchmarks, and an overall prediction error of 22% for the organic precursors for six active pharmaceutical ingredients.